Special Section on Video Surveillance and Transportation Imaging Applications

Video-based real-time on-street parking occupancy detection system

[+] Author Affiliations
Orhan Bulan

Xerox Research Center Webster, Webster, New York 14580

Robert P. Loce

Xerox Research Center Webster, Webster, New York 14580

Wencheng Wu

Xerox Research Center Webster, Webster, New York 14580

YaoRong Wang

Xerox Research Center Webster, Webster, New York 14580

Edgar A. Bernal

Xerox Research Center Webster, Webster, New York 14580

Zhigang Fan

Xerox Research Center Webster, Webster, New York 14580

J. Electron. Imaging. 22(4), 041109 (Aug 12, 2013). doi:10.1117/1.JEI.22.4.041109
History: Received April 15, 2013; Revised June 27, 2013; Accepted July 16, 2013
Text Size: A A A

Open Access Open Access

Abstract.  Urban parking management is receiving significant attention due to its potential to reduce traffic congestion, fuel consumption, and emissions. Real-time parking occupancy detection is a critical component of on-street parking management systems, where occupancy information is relayed to drivers via smart phone apps, radio, Internet, on-road signs, or global positioning system auxiliary signals. Video-based parking occupancy detection systems can provide a cost-effective solution to the sensing task while providing additional functionality for traffic law enforcement and surveillance. We present a video-based on-street parking occupancy detection system that can operate in real time. Our system accounts for the inherent challenges that exist in on-street parking settings, including illumination changes, rain, shadows, occlusions, and camera motion. Our method utilizes several components from video processing and computer vision for motion detection, background subtraction, and vehicle detection. We also present three traffic law enforcement applications: parking angle violation detection, parking boundary violation detection, and exclusion zone violation detection, which can be integrated into the parking occupancy cameras as a value-added option. Our experimental results show that the proposed parking occupancy detection method performs in real-time at 5frames/s and achieves better than 90% detection accuracy across several days of videos captured in a busy street block under various weather conditions such as sunny, cloudy, and rainy, among others.

Urban parking management is receiving significant attention due to its potential to reduce traffic congestion, fuel consumption, and emissions.1,2 Real-time parking occupancy detection is a critical component of on-street parking management systems, where occupancy information is relayed to drivers via smart phone apps, radio, Internet, on-road signs, or global positioning system auxiliary signals. Based on the studies performed in the central business districts of 11 cities on four continents, including New York, London, and San Francisco, it is estimated that 35% of the cars that make up the city’s traffic flow are cruising for parking. The traffic congestion created by vehicles searching for an available parking space, in turn, has significant impact on fuel consumption. Based on some studies, for example, it is estimated that half of the fuel consumed in San Francisco is due to vehicles searching for vacant spaces.3 It is also known that the majority of vehicles cruising in traffic searching for a parking space is looking for curb parking (i.e., on-street parking) because the average price for on-street parking is about 20% of the price of off-street parking, according to a study performed in 20 cities throughout the United States.1,2 Real-time on-street parking occupancy detection and broadcasting the available parking space are, therefore, essential to reduce traffic congestion and fuel consumption.

Sensors are required to gather real-time data from on-street parking areas. Among the many types of sensors, three types are currently deployed in automated parking occupancy detection systems: inductive loops, ultrasonic sensors, and magnetic in-ground sensors.4,5 Inductive loops and ultrasonic sensors are typically used in parking lot and garage settings where sensors are installed at the entrance and exit to count the vehicles entering/exiting the parking area and hence, are capable of estimating the parking occupancy rate in a parking lot or garage indirectly by measuring vehicle flow. These sensors are not applicable in on-street parking setting where the parking area is located along the traffic flow.

Magnetic in-ground (i.e., “puck-style”) sensors4 constitute another type of sensor technology currently used in on-street parking occupancy detection systems. These sensors are buried within the road in each parking stall and wirelessly communicate the vehicle detection information to interested parties. This particular sensor-based solution, however, has several disadvantages. There is a relatively high cost for in-ground sensor installation and maintenance. Also, there may be a need for more than one sensor per parking spot for more reliable detection. In addition, if a sensor requires replacement or other maintenance, there is likely a need to close one or more parking spots during the maintenance period, thus reducing parking efficiency. Sensor battery life is another concern; batteries typically last anywhere from 3 to 5 years. Furthermore, magnetic sensors perform poorly when there exists an inherent magnetic field on the street due to, for example, the existence of a nearby train station; they are also susceptible to temporary perturbations in the surrounding magnetic field due to, for instance, a large truck traveling along the street.6

Video-based sensing for monitoring on-street parking areas offers several advantages. One advantage is that one video camera can typically monitor and track several parking spots, whereas multiple magnetic sensors may be needed to reliably monitor one parking space. Another advantage is that maintenance of the video cameras is likely to be less disruptive than maintenance of in-ground sensors. Video cameras can also support other tasks such as traffic law enforcement and surveillance since these cameras capture a wider range of useful information including vehicle color, license plate, vehicle type, speed, etc. A video-based parking occupancy detection system can, therefore, provide a convenient, cost-effective solution to the sensing task while providing additional functionality for traffic law enforcement and surveillance.

In this article, we present, to the best of our knowledge, the first video-based on-street parking occupancy detection system that can operate in practical street settings with high detection performance in real time. Our system includes video cameras monitoring on-street parking areas that continuously capture video and transfer the image sequences to a central processing unit; the videos are then processed in order to estimate the on-street parking occupancy, which can then be shared publicly. Our method addresses several challenges an on-street parking setting poses to a vision-based system, namely shadows, reflections, illumination changes, occlusions, rain and other inclement weather conditions, and camera shake due to wind or vehicle-induced vibration. The method utilizes several components from video processing and computer vision for motion detection, background subtraction, and vehicle detection. In addition to the parking occupancy method, we also present several traffic law enforcement applications that can be enabled by on-street parking occupancy cameras. These additional functions can be integrated into parking occupancy cameras as a value-added option. Our experimental results performed on a busy street block across several days show the effectiveness of the method. Our occupancy detection method was also tested at various conditions to show the robustness of the method against typical challenges that exist in on-street parking management.

A key distinction of our method relative to a recently proposed parking occupancy determination method7 is that we operate on video, whereas, the method proposed in 7 operates on still images and does not address challenges specifically related to video processing (e.g., determining the frames for reliable vehicle detection). There are also other vehicle detection methods in the literature that operate on still images.8,9 These methods do not exploit temporal information in video leading to the need for more complex processing algorithms, which constitutes a severe challenge in implementing real-time applications.

The remainder of this paper is organized as follows. Our parking occupancy detection method is described in Sec. 2. Section 3 presents several traffic law enforcement applications that can be enabled by on-street parking occupancy cameras. Experimental results across several days of videos captured at different conditions are presented in Sec. 4. Section 5 concludes the article by summarizing key aspects of the video-based parking occupancy system.

An overview of the parking occupancy detection system is illustrated in Fig. 1. Our system consists of two phases: offline training and real-time operation. The offline phase defines a rectified region of interest (ROI) corresponding to the parking area in the scene and trains a support vector machine (SVM) classifier using the positive (vehicle) and negative (non-vehicle) samples extracted from the ROI of the videos acquired by the camera. The offline-trained SVM parameters are used in the initialization and verification/localization steps of the operational phase. After the initialization step, which will be discussed in more detail in Sec. 2.2, video frames are processed to detect candidate regions where a new vehicle may exist. Candidate regions are determined by foreground blobs in which no motion or occlusion is detected. The location of candidate regions in a frame is indicated by a binary mask as shown in Fig. 1. The regions indicated by pixels with nonzero value, or active pixels, in the mask capture changes caused by a parked vehicle as well as changes due to varying illumination, shadows, glare, reflections, and other objects. In order to detect parked vehicles and eliminate blobs caused by other factors, the candidate regions are verified and localized by applying the classifier trained in the offline phase in a sliding window search mode. Our method performs the computationally expensive sliding window search only in candidate regions, and only when necessary, thereby, enabling real-time performance.

Graphic Jump LocationF1 :

Overview of the video-based parking occupancy detection system.

Offline Operations: Defining ROI, Training for Vehicle Detection

Detecting/localizing parked vehicles in the ROI is required in the operational phase of our system. We adopt a computer vision approach to detect parked vehicles from still images, which requires an offline (training) phase for training a classifier.8,10,11Figure 2 illustrates an overview of the offline phase used in our system. An ROI is defined on the image plane based on the sizes of the image areas spanned by typical passenger cars in the parking region. The location of the parking region on the image plane depends on the geometric configuration of the camera set-up and is specified manually at camera set-up/installation. Once the ROI is defined on the image plane, the image/video frame is rotated so that the ROI is oriented along the horizontal direction, as shown in Fig. 2. The positive (vehicle) and negative (nonvehicle) samples are extracted from the rotated ROI. Extracting the training samples from the rotated ROI results in a reduction of the search space dimensionality considered in the sliding window search process, as will be explained in Sec. 2.2.

Graphic Jump LocationF2 :

Overview of the offline (training) phase.

After collecting a sufficient number of positive and negative samples, we construct our vehicle detection classifier using histograms of oriented gradients (HOG)11 as features and a linear SVM as our machine learning engine/classifier; linear SVMs are much more computationally efficient than the more sophisticated latent SVM used in 10, which is important to real-time performance for the later steps of verification and localization of vehicles within candidate regions, as described in Sec. 2.4. The computational efficiency and the performance of HOG have previously been successfully demonstrated for object detection from stationary images.11 We then perform a cross-validation test to select an appropriate feature dimension. In particular, we constrain the dimensionality of our HOG feature vector to N×2N×9, which means that the gradient image of a training sample is segmented into N×2N nonoverlapping blocks and 9-orientation-bin histograms of gradients in each block are computed as suggested in 11. Note that the number of blocks along the horizontal direction is twice the number of blocks N along the vertical direction; this aspect ratio is set based on the average aspect ratio of the positive samples. We perform the cross-validation test as follows:

For each N=1,2,,10,

  1. Randomly select 75% of positive and negative samples to train a linear SVM classifier.
  2. Evaluate the classification accuracy on the training set.
  3. Evaluate the classification accuracy on the remaining 25% of the samples (i.e., test set).
  4. Repeat the first three steps five times.

The classification accuracy across various feature sizes is shown in Fig. 3(a). The figure shows the average classification accuracy over five iterations on the test set with the error bar corresponding to the standard deviation. The classification accuracy on the training set is also shown in the figure. Note that the training error is 0 for all feature sizes except for N=1. The classification error on the test set, however, is significantly lower for N=7, 8, and 9. Among these three feature sizes, the number of support vectors is the smallest for N=7 as shown in Fig. 3(b). We, therefore, set the feature size to N=7 (i.e., 7×14×9).

Graphic Jump LocationF3 :

Cross-validation test for determining histograms of oriented gradients (HOG) feature dimension for our vehicle classifier: (a) number of classification errors versus feature dimension and (b) number of support vectors in the resulting classifier versus feature dimension.

Note that the offline phase includes defining an ROI on the image plane and training a classifier using the positive and negative samples obtained from each camera installed on site. Even though these operations require manual processing, this is a one-time effort and can be performed during camera installation/set-up. The field of view of a camera can be changed over time due to factors such as severe weather, vandalism, etc., which mandates reperforming the offline operations. We, however, observe that the outdoor cameras can be installed to be quite robust to weather conditions. Alternatively, a robust classifier can be trained to detect vehicles in any position and orientation12 on images to be used in cameras installed on different sites. While a robust classifier can eliminate the process of collecting training samples for each camera, the detection performance of these classifiers is significantly lower than our site-specific classifier.12 In a parking occupancy detection system, the primary concern is detection accuracy and extensibility/maintenance of video cameras is only a secondary concern. In this article, we therefore adopt a site-specific classifier to achieve high detection performance.

Initialization

Initialization estimates parking occupancy based on initial frames using stationary images. As in current state-of-the-art object detection methods,8,10,11 the vehicle detection process comprises the following steps: sliding window search, window classification, and nonmaximum suppression; this is illustrated in Fig. 4. Note that the sliding window search is only performed through the ROI carved out from the rotated image. The advantage of collecting training samples from the rotated image in Sec. 2.1 becomes evident in this context since it allows the sliding window search to be performed in a two-dimensional space (i.e., along the horizontal direction and for different window widths), therefore improving the processing time. This is in contrast to the more traditional sliding window approach which performs searches on a four-dimensional space, namely, along the horizontal and vertical directions and for different window widths and heights. Each search window is classified using the classifier trained in the offline phase. The classifier assigns a score for each search window indicating the estimated likelihood of the window containing a vehicle. A higher SVM score indicates a higher likelihood of the window containing a vehicle. The classification is performed by thresholding the calculated SVM score. This threshold is determined in the offline phase after training the classifier. In order to determine the threshold, we perform a sliding window search along the parking region in video frames acquired in the offline phase. The sliding window search assigns a score for each search window. The threshold is determined based on the scores of negative (nonvehicle) and positive (vehicle) search windows to achieve the best classification performance. In our experiments, we set this threshold from 0.7 to 0.8 based on the training images in the offline phase. Finally, a nonmaximum suppression method is applied to yield final locations of vehicles based on the highest scores among the overlapping windows containing positive results.13

Graphic Jump LocationF4 :

Vehicle detection in the first frame using the classifier and two-dimensional sliding window search in the region of interest (ROI).

Candidate Region Detection

We identify candidate regions where a parked vehicle exists in a video frame and restrict vehicle detection processing to these regions. Not all the frames in the video are relevant to parking occupancy estimation. For example, in most of the frames acquired during rush-hour, the parking region is typically occluded by vehicles traveling along the street in the flow of traffic. Sliding window search methods applied to these frames would lead to false detections and missed vehicles. This is illustrated in Fig. 5(a), where a truck traversing the scene occludes the parking region. Another example of a frame on which vehicle detection methods may be unreliable is illustrated in Fig. 5(b), which shows a scene where a vehicle is in the process of parking. In summary, we select candidate regions among locations in the ROI where a foreground blob is detected and discard the foreground blobs in which motion/occlusion is detected. In the remainder of this section, we first describe a background subtraction method to determine foreground blobs in the ROI and then describe motion and occlusion detection methods that allow us to discard foreground blobs in which motion/occlusion is detected when identifying candidate regions for vehicle detection.

Graphic Jump LocationF5 :

Examples of unreliable frames for estimating parking occupancy illustrating (a) an occlusion and (b) a vehicle in the process of parking.

Background estimation and subtraction

Foreground objects in the scene can be identified by using a background subtraction method whereby a background estimate is subtracted from the current frame. The subtraction is performed as a pixel-wise distance metric in a predetermined intensity or color space. Pixels for which the computed distance is larger than a predetermined threshold are classified as foreground pixels. Additional morphological operations are applied that remove outliers and fill in holes resulting in a binary mask identifying the location of foreground areas. There are several techniques for background estimation including those based on Gaussian mixture models,14 eigenbackgrounds that use principal component analysis,15 and computation of running averages16 that gradually update the background as new frames are read.

Due to its simplicity and efficiency, we adopt a running average algorithm for background estimation and modify the algorithm for our application. In our method, the background is initialized as the first frame in the video sequence and gradually updated via a running average, autoregressive process as new frames are read. The background update is performed as Display Formula

Bt+1=f×Ft+1+(1f)×Bt,(1)
where Bt is the background at time t, Ft+1 is the frame at time t+1, and f is the update factor.

Although the use of running average algorithms in video processing is not new, we incorporate a novel selection process for the updating factor f for each pixel, where the selection is particularly suited for determining parked vehicles that may stay in the scene for a long time. Display Formula

f={0ifmotion/occlusion is detected or if there is a parked vehiclecotherwise

We set the update factor f for a pixel indexed by (i,j) to 0 if motion/occlusion is detected at that pixel location or if there is a parked vehicle at that position; background pixels do not get updated for these locations under these conditions so that a foreground blob is detected as long as a parked/occluding/moving vehicle stays in the scene. Note that pixels labeled as vehicles at initialization fall under this category since they correspond to regions associated with parked vehicles. The update rate c determines how fast background is updated and can be set according to frame rate and other specifications in an application. A reasonable c value, for example, is 0.05 for a 30 fps video.

Once the background is estimated at a given time, the foreground objects in the parking area are detected by subtracting the current frame from the estimated background and applying thresholding and morphological operations on the absolute difference image. We set the threshold to 10 in a gray-scale space to enable detection of changes in the parking region when the acquired video is low contrast due to weather conditions. The candidate regions are identified among these foreground blobs, in which no motion/occlusion is detected.

Motion detection

Motion detected in the ROI indicates a moving vehicle in the parking region and motion detected outside ROI on the roadway refers to an occluding vehicle traversing the scene. In either case, the motion detected region is unreliable for detecting a parked vehicle in the parking area. We therefore exclude detected motion blobs from candidate regions in the vehicle detection step. Motion detection can be performed by calculating motion vectors between two adjacent frames in a video sequence using a pixel-level optical flow method17,18 or a block-matching algorithm.19 These techniques calculate motion vectors for each pixel/block and therefore are typically computationally expensive. Temporal difference methods, on the other hand, are simpler and operate by subtracting adjacent frames followed by thresholding, yielding a binary mask indicating a region of change.20,21 We use a double-difference algorithm followed by morphological filtering to detect the objects in motion. Note that any number of vehicles in motion in the parking region can be detected using the double difference algorithm and the regions identified by the motion detection algorithm are excluded from candidate regions.

Occlusion detection

A foreground blob can correspond to a parked vehicle in the ROI as well as to vehicles occluding the parking area. We eliminate foreground blob regions due to an occluding vehicle and exclude them from the candidate regions since they are not reliable for vehicle detection. While this aids robustness, it also helps reduce computation involved in verification and localization steps. Note that occlusion can also be caused by stationary objects, in which case occlusion determination via motion detection is insufficient. For example, an occluding vehicle in a traffic lane can stop at a traffic light or due to traffic congestion. Our algorithm detects the occurrence of an occlusion event when a detected foreground blob is partially inside the parking region and partially outside the region, as illustrated in Fig. 6. Note that a properly parked vehicle is entirely inside the ROI and hence, is not deemed to be an occluder. Also a vehicle parked beyond the parking stall boundaries on the roadway side is not detected as an occlusion. This is because the majority of the pixels in the detected blob in this case are inside the ROI, whereas the majority of the detected blob is outside the ROI in the case of an occlusion. A suitable threshold can be set for the area of a detected blob inside and outside the ROI to separate occlusion and out-of-box parking. This threshold can be set based on camera configuration, geometry, and parameters (e.g., camera resolution).

Graphic Jump LocationF6 :

Occlusion is detected when a foreground blob is partially inside the parking area of interest and partially outside the parking region.

Verification and Localization

Candidate regions do not always indicate the presence of a parked vehicle in the scene. We perform a verification procedure to distinguish vehicle blobs from nonvehicle blobs and eliminate the latter. Figure 7 illustrates regions detected by changes in foreground in several situations. Figure 7(a)7(c) shows a blob detected due to a parked vehicle along the street. Figure 7(d)7(f) illustrates a candidate region produced by the reflection of headlights on the road and does not contain a vehicle. Figure 7(g)7(i) shows another situation where the detected blob contains a parked vehicle as well as its shadow. Note that the detected blob corresponds strictly to a parked vehicle only in the first case.

Graphic Jump LocationF7 :

Candidate blobs detected due to changes produced by a parked vehicle (a–c), headlights (d–f), and a parked vehicle and its shadow (g–i).

When a blob large enough to be considered a potential vehicle is detected in the parking area, we test to verify that the detected blob indeed corresponds to a vehicle. Our test is based on a machine learning approach using the classifier trained in the offline phase (Sec. 2.1). The classifier assigns a score for the candidate region indicating a decision that the detected blob corresponds to a vehicle based on the calculated features. Note that this is not a sliding window search but rather a one-shot classification and hence, is computationally simple.

Although our verification procedure identifies a blob that strictly corresponds to a parked vehicle, such as in Fig. 7(c), it does not guarantee to identify a candidate region when it contains a vehicle and its shadow, as in Fig. 7(i). When features are extracted from the detected region containing a significant shadow and input to the classifier, it identifies the blob as nonvehicle, causing a missed detection. We perform a sliding window search around the neighborhood of the candidate region to locate parked vehicles when the candidate blob is identified as nonvehicle by our verification procedure. The search is performed across different window sizes in the local neighborhood to locate the vehicle within the candidate region.

Alternatively, a shadow suppression technique can be applied to eliminate portions of the blob that correspond to shadow areas. This can be performed by methods similar to that proposed in 22, which classifies regions whose chrominance is similar to the background but possesses lower luminance. Once the shadow is suppressed, the remaining portion of the blob can be passed to our verification procedure to identify the vehicle.

Note that our verification and localization module are complementary to other system elements and fix the errors caused by limitations of the preceding modules such as motion/occlusion detection, background subtraction, candidate region determination, etc. For this reason, we adopt simple but efficient algorithms for these tasks to reduce the computational load in the system. Alternatively, complicated and deeper algorithms can be developed for occlusion/motion detection, background subtraction, etc., but these algorithms typically come at a cost of higher computational requirements, which in turn impacts achieving real-time performance in the full system.

In many cities, on-street parking is regulated with substantial fines for violators.23,24 In 2009, for example, 10,662,000 tickets were issued for parking violations in New York City, generating $600 million in revenue.25 Traditionally, these regulations are manually enforced by authorized personnel or law enforcement officers. Parking occupancy cameras have the capability to perform their occupancy/vacancy function as well as automate the parking law enforcement process. Below, we present several law enforcement applications that can be performed by a parking occupancy camera system.

Parking Angle Violation Detection

In on-street parking, parking/standing at a substantial angle to the curb is prohibited by traffic laws.23 Similarly, parking parallel to the curb is prohibited in areas where parking at an angle is authorized.23

Our video-based parking angle violation detection method builds on the vehicle detection framework of Sec. 2. Our method estimates the angle of a parked vehicle relative to the curb by using one of the following algorithms:

  1. Identifying the longest line in the image blob corresponding to a detected vehicle and calculating the angle of the identified line.
  2. Fitting an ellipse to the detected blob and calculating the angle of the major axis.

The first algorithm extracts edges of the detected vehicle and searches for the longest line within the blob. The angle of this line is a good estimate of the angle at which the vehicle is parked because the longest line detected within the blob is typically the line that separates the windows from the body or the body from the road provided that there is enough contrast between the window and body and that the vehicle is not parked at an angle close to 90 deg (right angle) to the curb. The longest line can be found using an algorithm such as a Hough transform.26 This is illustrated in Fig. 8(c) which shows the binary mask containing the detected blob, and Fig. 8(d) shows the longest line detected within the blob. The parking angle of the vehicle can be calculated as θ=atan[(y2y1)/(x2x1)] where θ is the angle of the parked vehicle with respect to image plane and (x1,y1) and (x2,y2) are the coordinate locations of the start and end points of the detected line, respectively.

Graphic Jump LocationF8 :

Parking angle detection is performed by finding the longest line in the detected vehicle: (a) current frame, (b) estimated background, (c) binary mask illustrating the location of the detected vehicle, and (d) longest line (green) identified in the image region corresponding to the detected vehicle.

A second algorithm that can be used to estimate the angle of a parked vehicle fits an ellipse to the detected blob and determines the angle of major axis. An ellipse can be fitted to a blob using one of the following approaches:

  • Perform regression to fit an ellipse to the points at the boundary of the detected blob.
  • Find the orthogonal pair of segments with the highest length ratio among all line segments that pass through the centroid of the convex hull containing the detected blob.

In the first approach, the regression minimizes the sum of the squared distances from the perimeter of the ellipse to the data points along a radial line extending from the center of the ellipse to each data point.27

In the second approach, the convex hull containing the blob is computed. The centroid of the resulting convex blob (x¯,y¯) is calculated according to Display Formula

x¯=xyI(x,y)xxyI(x,y),y¯=xyI(x,y)yxyI(x,y),
where (x¯,y¯) are the coordinates of the blob centroid, (x,y) are the coordinates of the pixels in the blob, and I(x,y) is the binary image values in the detected blob. The major and minor axes of the ellipse can be determined among all line pairs that pass through the centroid of the blob and that are perpendicular to each other. Among the perpendicular line pairs passing through the centroid, the minor and major axes are selected as the pair with the highest length ratio (ratio of longer axis length to shorter axis length), as illustrated in Fig. 9. The blue circle inside the detected blob represents the centroid. The major and minor axes in the figure are shown as the blue perpendicular lines having the highest length ratio. The parking angle of the vehicle can be calculated from the orientation of the major axis.

Graphic Jump LocationF9 :

Fitting an ellipse to a blob. The minor and major axes of the ellipse are selected as the line segments that have the highest length ratio among all the perpendicular line segment pairs passing through the centroid.

To detect a violation, the absolute difference between the parking angle and curb angle is compared with a predetermined threshold T0, set according to the relevant legislation. If the difference is larger than T0, a signal can be transmitted to authorities so they may take action, such as issuing a ticket.

Parking Boundary Violation Detection

On-street parking areas are typically divided into stalls marked by painted lines (see Fig. 10). Each stall is associated with a parking meter to charge a vehicle according to the time it stays in the stall. Parking beyond the stall markings is prohibited by regulations in several cities, with an exception when a vehicle is too large to fit in that marked parking space.23

Graphic Jump LocationF10 :

Stall parking. The parking space for each vehicle is separated by lines painted on the road surface, shown as dashed lines in the figure.

Here, we present a video-based method to detect vehicles parked beyond the boundaries of a parking stall. Once a vehicle is detected in a parking area, the method classifies the detected vehicle into one of two classes: a short (vehicles that would fit in a single stall, e.g., cars, vans, sport utility vehicles); long (vehicles that would not fit in a single stall, e.g., trucks). Vehicle classification can be simply implemented in the following ways.

  • Use of geometric attributes of a detected blob (e.g., area, length, height, width, eccentricity, combinations thereof, etc.).
  • Estimating physical length of the detected vehicle using a calibration technique.

Note that we employ existing more elaborate techniques for video-based vehicle classification that can separate vehicles into many classes,2832 but we opt for simpler techniques that are consistent with parking regulations. A straightforward method is based on the area of the detected blob for a parked vehicle. The area on the image plane of long vehicles is typically much larger than the area of short vehicles, as shown in Fig. 11. The blob area is compared with a predetermined threshold T1 to distinguish long vehicles from short vehicles, where the threshold depends on camera configuration, geometry, and parameters (e.g., camera resolution) and can be set at the camera installation/set-up before initializing the algorithm. For a camera with a resolution of 640×512 and with a field of view as shown in Fig. 11, for example, we set the threshold to 104 to distinguish large-size vehicles from the passenger size vehicles. We, however, observe that there is latitude for selecting this threshold for the camera configuration shown in Fig. 11.

Graphic Jump LocationF11 :

Vehicle classification based on the area of detected blobs for parked vehicles: (a) current frame and (b) detected blobs for a long (truck) and short (car) vehicle.

Another way to perform vehicle classification is by estimating the physical length d of a detected vehicle. This can be achieved by first determining the start and end points of the detected blob for a parked vehicle. A line is drawn through each of the start and end points perpendicular to the street direction on the image plane. The line that joins the start and end points is projected onto the line perpendicular to the lines that pass through the start and end points. The vehicle length is estimated as the length of the projected line, as illustrated in Fig. 12. The physical length of the projected line can be estimated through a calibration that maps pixel coordinates to real-world coordinates.33 The estimated length d is compared with a predefined threshold T2. The threshold T2 is set to a typical stall length of 20 ft to identify vehicles that are longer than a typical stall, which are allowed to park beyond parking boundaries.

Graphic Jump LocationF12 :

Vehicle classification by estimating the length of a vehicle.

A violation is detected when a vehicle is classified as a short vehicle and parked beyond the parking boundaries. This can be performed by calculating the area or distance metric of the parked vehicle out of the boundary and compared with a predetermined threshold T3. In case of detecting a violation using a distance metric, for example, T3 can be set to 5 ft, approximately a quarter of a typical parking stall length, to identify short vehicles that violate the parking boundary regulation more than quarter of a typical stall length. This threshold can be adjusted according to the relevant legislation.

Exclusion Zone Violation Detection

A common parking regulation restricts/prohibits parking/standing/stopping in exclusion zones (e.g., within a specified vicinity of a fire hydrant, private road, garbage container, bus stop, sidewalk/crosswalk, stop sign, on roadway side of a vehicle parked at the curb, etc.). In New York City, for example, more than 30% of the parking violation codes regulate parking in exclusion zones.23 In 2005, infractions of exclusion zone regulations were among the most common three NYC parking violations, leading to over 1.5 million issued tickets.34

The video-based vehicle detection method described in Sec. 2 can be used to detect vehicles parked in exclusion zones. Figure 13, for example, shows three exclusion zones delimited by red boxes and manually specified on an acquired video frame. The specified regions in Fig. 13(a) and 13(b) show exclusion zones in front of a private driveway and a fire hydrant, respectively. The exclusion zone specified in Fig. 13(c) corresponds to a portion of a road that is adjacent to an on-street parking area where double parking can occur. If the exclusion zone is a region where stopping/standing/parking is prohibited at all times, a signal is transmitted immediately to authorized entities when a parked vehicle is detected.

Graphic Jump LocationF13 :

Exclusion zones (a) in front of a private driveway, (b) in front of a fire hydrant, and (c) on a road adjacent to an on-street parking area.

Performance of the On-Street Parking Occupancy Detection Method

We evaluated the performance of our parking occupancy detection method on videos captured on a city block with two-way traffic. On-street parking was allowed on each side of the block, having several challenges that exist in a typical on-street parking setting such as trees, shadows, occlusions, reflections, people moving/standing in the parking area, etc. The parking regions along the block were designated for multispace parking, where there were no predefined boundaries marked by lines on the road. We installed three Vivotek IP8352 surveillance cameras in the block, each monitoring a different portion of the parking area along the street as shown in Fig. 14. Cameras 1 and 3 were installed on the same side of the street at the two ends of the block, and camera 2 was installed on the other side monitoring the parking area across the region monitored by cameras 1 and 3. We defined the ROI in the videos captured by cameras 1 and 3 such that half of the parking area in the scene is covered by camera 1 and the other half is covered by camera 3. We recorded several days of video from each camera at various weather conditions such as sunny, cloudy, rainy etc., as illustrated in Fig. 15. The captured videos had a resolution of 640×512 and a frame rate of 5 fps. The videos were captured from 7:30 am to 7:00 pm.

Graphic Jump LocationF14 :

Sample video frames illustrating the field of view of the cameras in the block (a) Camera 1, (b) Camera 2, and (c) Camera 3.

Graphic Jump LocationF15 :

Sample video frames illustrating different weather conditions at which test videos were captured.

In the offline phase of our method, we trained a linear SVM classifier for each camera as described in Sec. 2.1. For this purpose, we recorded several days (i.e., 5 to 10 days) of video and collected around 600 positive and 1200 negative samples from the recorded videos for each camera. The samples were manually cropped from the recorded videos without any preprocessing. The positive samples include different vehicles parked along the block and were collected from 300 different frames captured by each camera. Negative samples were cropped randomly from the parking region in the recorded videos for each camera. We used two-thirds of the positive and negative samples for training and the remaining one-third for testing. The number of support vectors for the trained classifiers was 100, 81, and 74 for cameras 1, 2, and 3, respectively, which correspond to 6% to 8% of the total number of training samples for a camera. The trained classifiers for each camera achieved zero training error and at most 3% error in the test set. Since our classifiers were trained site specific, they do not perform well on videos captured by cameras installed in different geometry and configuration.

After training the classifiers in the offline phase, we implemented the video-based parking occupancy detection method in MATLAB and tested its performance in the operational phase on a 2.2 GHz i7 machine with 16 GB memory. Our method was able to process the videos faster than 5 fps and hence, achieve the real-time processing. The performance is evaluated in terms of detection accuracy calculated as Display Formula

Detection accuracy=1False alarms+missed detectionsNumber of parking events,(2)
where the number of parking events is determined by the number of incidents that a new vehicle parks or a parked vehicle leaves the parking region. Table 1 shows the detection accuracy across 5 days of videos captured by each camera. These 5 days of videos are different from the videos utilized for collecting the samples to train the classifier. The number of parking events was changing from 120 to 200 depending on the day and parking location. As shown in the table, the average detection accuracy for each camera is over 91% which is higher than the detection accuracy required by in-ground sensors in San Francisco.35 Note that the detection performance is slightly lower for the videos captured by camera 3 because the camera has an oblique view and vehicles parked close to middle of the block sometimes get occluded by other parked vehicles, causing missed detections. Another reason for missed detections is due to black cars parking under tree shadows on a sunny day, which could not get detected by our background subtraction method. This error can be eliminated by adjusting the threshold for background substraction but at a cost of performing sliding window search more often, which increases the processing time.

Table Grahic Jump Location
Table 1Detection accuracy of the video-based parking occupancy detection method across 5 days of video captured from three different vantage points.

We have also calculated probability of false positive (FP) and negative (FN) for our parking occupancy detection algorithm where the probability of FP and FN are calculated as Display Formula

Probability of false positive=Number of false alarmsNumber of parking events(3)
Display Formula
Probability of false negative=Number of missed detectionsNumber of parking events.(4)

From our experiments across several days of videos captured by our three cameras, we calculated the probability of false positive as 2.92% and the probability of false negative as 3.21% for our parking occupancy detection system. Note that the classifier parameters can be changed to tune classifiers toward reducing number of false positives at a cost of increasing number of false negatives or vice versa.

Experimental Examples of the Law Enforcement Methods
Parking angle violation detection

During our parking occupancy detection experiments, we did not observe any parking angle violation in the scene. We therefore staged this violation in a video captured on a local road. The video was taken with a commercially available Vivotek IP8352 surveillance camera. The captured video had a frame rate of 5 fps and a resolution of 300×256pixels.

In our experiment, we consider three different parking angle scenarios. The first setting corresponds to lawful parking where the vehicle is parked along the street direction and parallel to the curb. In the second and third settings, the vehicle is parked at an angle with respect to the curb as shown in Fig. 16(b) and 16(c).

Graphic Jump LocationF16 :

Sample video frames illustrating different parking scenarios.

We calculated the curb angle manually as 2 deg before initializing the algorithm. The parked vehicles were then detected using the algorithm described in Sec. 2 and the angle of the parked vehicles were estimated by fitting ellipses to the detected blobs, as shown in Fig. 16(d)16(f). The orientations of the ellipses were calculated as 4, 5, and 18 deg, respectively, by our algorithm. We also computed the angle of the parked vehicles by manually calculating the angle of the line on the side of the vehicle from high-resolution versions of the frames. The manually calculated angles were 3, 6, and 18 deg, respectively, which were closely estimated by our algorithm. As expected, the difference between the curb angle and the parking angle was the smallest for the lawful case, which was identified by setting a suitable threshold.

Parking Boundary Violation Detection

We tested our algorithm for detecting parking boundary violations on a video sequence captured on a road with stall parking, where each stall was marked by lines on the road. A 45-min video was captured with a Vivotek IP8352 surveillance camera monitoring the parking area. The captured video had a frame rate of 5 fps and a resolution of 300×256pixels.

We manually determined the boundaries of the parking stalls within the camera view. We then initialized the algorithm by setting the first frame as the background. The algorithm described in Sec. 2 updated the background gradually and detected the parked vehicles in the parking region along the street. When we started the video capture, the silver car was in the scene and so it was detected in the initialization step of the algorithm.

Figure 17 shows the results of our method for detecting vehicles that violate the parking boundary regulation. In the figure, the images in the first column show the current frame and those in the second column show the detected stationary vehicles in the parking area. The vehicle that violates the regulation is indicated by a red dashed box and the vehicle that is classified as a long vehicle is indicated by a white dashed box. Note that the vehicle in the first row is classified as a short vehicle and it parks beyond the boundaries of a parking stall and hence, is detected as a violator. The delivery truck is, however, classified as a long vehicle and does not violate the law even though it parks beyond the boundaries of a parking stall and occupies more than one stall. All the other vehicles in the video sequence are classified as short vehicles and are parked within the parking boundaries.

Graphic Jump LocationF17 :

Parking boundary violation detection results. The vehicle inside the red box violates the law. The truck inside the white box is classified as a long vehicle and does not violate the law even though it passes over the boundaries of a parking stall.

We presented a vision-based real-time parking occupancy detection system that addresses several challenges that exist in on-street parking. The computational cost of object detection methods is mitigated by using simple but efficient video processing algorithms for background subtraction, motion detection, and occlusion detection. Modifying the learning parameter in background construction based on running averages is particularly well suited for detecting candidate regions for a parked vehicle. Occlusion detected by the position of a foreground blob with respect to a parking region is used to eliminate unreliable frames/regions for vehicle detection. A verification procedure based on a machine learning approach can be effectively used for refinement of candidate regions and localizing parked vehicles within these regions.

In the future, we conjecture that video cameras will replace in-ground sensors for real-time parking occupancy determination due to their advantages in terms of installation/maintenance costs and extensibility to other law enforcement applications. This article paves the way toward developing a video-based real-time parking occupancy detection system that can be deployed in large scale. The law enforcement applications presented in this article are a sample of applications that can be offered as a value-added option to parking occupancy cameras. Since several of these violations occur rarely in real life, we could only show limited experimental examples. These methods, however, need to be evaluated and tested across larger set of videos captured from real life, which is an ongoing work that we have undertaken. Also note that our parking occupancy method includes an offline phase that requires manual intervention to collect training samples. Even though this manual intervention is performed only in camera installation/set-up, it can be cumbersome in large-scale deployment of parking occupancy cameras. Automating the offline processes is, therefore, desirable to lower the installation/maintenance cost for parking occupancy cameras, which is a part of our future work.

Shoup  D. C., The High Cost of Free Parking. , Vol. 7,  Planners Press, American Planning Association ,  Chicago  (2005).
Shoup  D. C., “Cruising for parking,” Transport Policy. 13, (6 ), 479 –486 (2006). 0967-070X CrossRef
Loce  R. P., Saber  E., “Fuel consumption in San Francisco,” http://spie.org/x1816.xml ( March 2013).
Wolff  J. et al., “Parking monitor system based on magnetic field sensors,” in  IEEE Intell. Transport. Syst. Conf. , pp. 1275 –1279,  IEEE ,  Toronto, Canada  (2006).
Mimbela  L. E. Y., Klein  L. A., “Summary of vehicle detection and surveillance technologies used in intelligent transportation systems,” http://www.fhwa.dot.gov/ohim/tvtw/vdstits.pdf ( August 2013).
Shaheen  S., Rodier  J. C., Eaken  M. A., “Smart parking management field test: a bay area rapid transit (BART) district parking demonstration,” Technical Report by Institute of Transportation Studies, UC Davis (2005).
Deruytter  M., Anckaert  K., “Video-based parking occupancy detection,” Proc. SPIE. 8663, , 86630O  (2013).CrossRef
Agarwal  S., Awan  A., Roth  D., “Learning to detect objects in images via a sparse, part-based representation,” IEEE Trans. Pattern Anal. Mach. Intell.. 26, (11 ), 1475 –1490 (2004). 0162-8828 CrossRef
Tsai  L.-W., Hsieh  J.-W., Fan  K.-C., “Vehicle detection using normalized color and edge map,” IEEE Trans. Image Process.. 16, (3 ), 850 –864 (2007). 1057-7149 CrossRef
Felzenszwalb  P. F. et al., “Object detection with discriminatively trained part-based models,” IEEE Trans. Pattern Anal. Mach. Intell.. 32, (9 ), 1627 –1645 (2010). 0162-8828 CrossRef
Dalal  N., Triggs  B., “Histograms of oriented gradients for human detection,” in  Proc. IEEE Computer Society Conf. Computer Vision and Pattern Recognition , Vol. 1, pp. 886 –893,  IEEE ,  San Diego, CA  (2005).
Felzenszwalb  P. F. et al., “Object detection with discriminatively trained part-based models,” IEEE Trans. Pattern Anal. Mach. Intell.. 32, (9 ), 1627 –1645 (2010). 0162-8828 CrossRef
Neubeck  A., Van Gool  L., “Efficient non-maximum suppression,” in  Int. Conf. Pattern Recognition , Vol. 3, pp. 850 –855,  IEEE ,  Hong Kong  (2006).
Stauffer  C., Grimson  W. E. L., “Adaptive background mixture models for real-time tracking,” in  Proc. IEEE Computer Society Conf. Computer Vision and Pattern Recognition , Vol. 2, pp. 252 ,  IEEE ,  Fort Collins, CO  (1999).
Oliver  N. M., Rosario  B., Pentland  A. P., “A bayesian computer vision system for modeling human interactions,” IEEE Trans. Pattern Anal. Mach. Intell.. 22, (8 ), 831 –843 (2000). 0162-8828 CrossRef
Lo  B., Velastin  S., “Automatic congestion detection system for underground platforms,” in  Proc. Int. Symp. Intelligent Multimedia, Video and Speech Processing , pp. 158 –161,  IEEE ,  Kowloon Shangri-La, Hong Kong  (2001).
Horn  B., Schunck  B., “Determining optical flow,” Artif. Intell.. 17, (1–3 ), 185 –203 (1981). 0004-3702 CrossRef
Lucas  B. et al., “An iterative image registration technique with an application to stereo vision,” in  Proc. Int. Joint Conf. Artificial Intelligence , Vol. 3, pp. 674 –679,  AAAI ,  Vancouver, Canada  (1981).
Huang  Y., Zhuang  X., “Motion-partitioned adaptive block matching for video compression,” in  Proc. IEEE Int. Conf. Image Processing , Vol. 1, pp. 554 –557,  IEEE ,  Washington, DC  (1995).
Makarov  A., Vesin  J.-M., Kunt  M., “Intrusion detection using extraction of moving edges,” in  Proc. 12th IAPR Int. Conf. Pattern Recognition , Vol. 1, pp. 804 –807,  IEEE ,  Jerusalem, Israel  (1994).
Paragios  N., Tziritas  C., “Detection and location of moving objects using deterministic relaxation algorithms,” in  Proc. 13th Int. Conf. Pattern Recognition , Vol. 1, pp. 201 –205,  IEEE ,  Vienna, Austria  (1996).
Cucchiara  R. et al., “Improving shadow suppression in moving object detection with HSV color information,” in  Proc. IEEE Intelligent Transportation Systems Conf. , pp. 334 –339,  IEEE ,  Oakland, CA  (2001).
NYC Department of Transportation, “Traffic violation codes in NYC,” http://www.nyc.gov/html/dof/html/parking/violation_codes.shtml ( January 2013).
Office of Documents and Administrative Issuance, Washington, DC, http://www.dcregs.dc.gov/Search/FullTextSearch.aspx?SearchType=DCMR&KeyValue=on+street+parking ( March 2013).
NYC Department of Transportation, “Parking tickets in NYC,” http://parkitnyc.com/parking-nyc/nyc-parking-tickets/ ( March 2013).
Illingworth  J., Kittler  J., “A survey of the Hough transform,” Comput. Vision Graphics Image Process.. 44, (1 ), 87 –116 (1988). 0734-189X CrossRef
Seber  G. A., Lee  A. J., Linear Regression Analysis. , Vol. 936,  John Willey & Sons ,  New York  (2012).
Chao  T.-H., Lau  B., Park  Y., “Vehicle detection and classification in shadowy traffic images using wavelets and neural networks,” Proc. SPIE. 2902, , 136 –147 (1997). 0277-786X CrossRef
Unzueta  L. et al., “Adaptive multicue background subtraction for robust vehicle counting and classification,” IEEE Trans. Intell. Transport. Syst.. 13, (2 ), 527 –540 (2012). 1524-9050 CrossRef
Morris  B., Trivedi  M., “Improved vehicle classification in long traffic video by cooperating tracker and classifier modules,” in  Proc. IEEE Int. Conf. Video and Signal Based Surveillance , p. 9 ,  IEEE  (2009).
Gupte  S. et al., “Detection and classification of vehicles,” IEEE Trans. Intell. Transport. Syst.. 3, (1 ), 37 –47 (2002). 1524-9050 CrossRef
Avery  R., Wang  Y., Scott Rutherford  G., “Length-based vehicle classification using images from uncalibrated video cameras,” in  Proc. IEEE Conf. Intelligent Transportation Systems , pp. 737 –742,  IEEE  (2004).
Zhang  Z., “A flexible new technique for camera calibration,” IEEE Trans. Pattern Anal. Mach. Intell.. 22, (11 ), 1330 –1334 (2000). 0162-8828 CrossRef
NYC Department of Transportation, “Exclusion zone parking tickets,” http://parkitnyc.com/parking-nyc/nyc-parking-tickets/ ( March 2013).
San Francisco Municipal Transportation Authority, “San Francisco parking occupancy detection performance requirement,” http://sfpark.org/wp-content/uploads/2011/09/SFpark_SensorPerformance_v01.pdf ( March 2013).

Grahic Jump LocationImage not available.

Orhan Bulan received a BS degree with high honors in electrical and electronics engineering from Bilkent University, Ankara, Turkey, in 2006, and MS and PhD degrees in electrical and computer engineering from University of Rochester, NY, in 2007, and 2012, respectively. He is currently a postdoctoral fellow in the Xerox Research Center Webster, Webster, NY. He was with Xerox during the summers of 2009, 2010, and 2011 as a research intern. He is the recipient of the best student paper award at the 2008 Western New York Image Processing Workshop organized by the Rochester Chapter of the IEEE Signal Processing Society. His recent research interests include signal/image processing, video processing, computer vision, and machine learning. He has 4 issued patents and over 10 pending patent applications in these areas.

Grahic Jump LocationImage not available.

Robert P. Loce is a research fellow and technical manager in the Xerox Research Center Webster, Webster, NY. He joined Xerox in 1981 with an associate degree in optical engineering technology from Monroe Community College. While working in optical and imaging technology and research departments, he received a BS in photographic science (RIT 1985), MS in optical engineering (UR 1987), PhD in imaging science (RIT 1993) and passed the US patent bar in 2002. A significant portion of his earlier career was devoted to development of image processing methods for color electronic printing. His current research activities involve leading an organization and projects into new video processing and computer vision technologies that are relevant to transportation, retail, and healthcare. He has publications and many patents in the areas of digital image processing, image enhancement, imaging systems, and optics. He is a fellow of SPIE and a senior member of IEEE. His publications include a book on enhancement and restoration of digital documents, and book chapters on digital halftoning and digital document processing. He is currently an associate editor for Journal of Electronic Imaging, and has been an associate editor for Real-Time Imaging, and IEEE Transactions on Image Processing.

Grahic Jump LocationImage not available.

Wencheng Wu is a principal scientist at the Xerox Research Center Webster, Webster, NY. He joined Xerox in 2000 with a PhD degree in electrical engineering from Purdue University, in West Lafayette, IN. His earlier career was focused on the areas of image quality metric developments, printer and sensor characterizations, image simulation and color modeling, color consistency measurement, and image processing algorithms for defect detection. His current research activities include computer vision, video processing, and video analytics for transportation applications. He has multiple papers and patents in areas related to his current and past research interests. He is a senior member of the Institute of Electrical and Electronics Engineers (IEEE) and a member of the Society of Imaging Science and Technology. He is also a frequent reviewer of IEEE Transactions on Image Processing, the Journal of Electronic Imaging and the Journal of Imaging Science and Technology.

Grahic Jump LocationImage not available.

YaoRong Wang is a research scientist at the Xerox Research Center Webster, Webster, NY, and received his PhD in physics in 1986 from Purdue University of West Lafayette, Indiana. His PhD thesis is on thermal and electronic properties of metals. After joining Xerox, he has worked on theory of semiconductor surface reconstruction, theory of ferromagnetic and antiferromagnetic materials, and theory of high-temperature superconductivity. Later on, he has worked on process control of xerographic printers, MEMS optical design, imaging and color management, intelligent sensing systems, and video based parking technology. He has 71 publications in variety of journals including Physical Review and Physical Review Letters, and holds 46 patents with 27 additional pending.

Grahic Jump LocationImage not available.

Edgar A. Bernal is a senior research scientist at the Xerox Research Center Webster, Webster, NY. He joined Xerox in 2006 with MSc and PhD degrees in electrical engineering from Purdue University, in West Lafayette, IN. His earlier career was focused on the areas of image processing, halftoning, image perception, watermarking, and color theory. His current research activities include computer vision, video compression, video-based object tracking, machine learning for financial data analytics and the application of novel sensing technologies to healthcare and transportation. He has multiple papers and patents in areas related to his current and past research interests. He is a senior member of the Institute of Electrical and Electronics Engineers (IEEE), and serves as the vice-chair of the Rochester, NY, chapter of the IEEE Signal Processing Society. He also serves as an adjunct faculty member at the Rochester Institute of Technology Center for Imaging Science and is a frequent reviewer of IEEE Transactions on Image Processing, the Journal of Electronic Imaging and the Journal of Imaging Science and Technology.

Grahic Jump LocationImage not available.

Zhigang Fan received his MS and PhD degrees in electrical engineering from the University of Rhode Island, Kingston, RI, in 1986 and 1988, respectively. He joined Xerox Corporation in 1988, where he is currently a principal scientist in Xerox Research Center Webster, Webster, NY. His research interests include various aspects of image processing and recognition. He has authored and coauthored more than 90 technical papers, as well as over 180 patents and pending applications. He is a fellow of SPIE and IS&T.

© The Authors. Published by SPIE under a Creative Commons Attribution 3.0 Unported License. Distribution or reproduction of this work in whole or in part requires full attribution of the original publication, including its DOI.

Citation

Orhan Bulan ; Robert P. Loce ; Wencheng Wu ; YaoRong Wang ; Edgar A. Bernal, et al.
"Video-based real-time on-street parking occupancy detection system", J. Electron. Imaging. 22(4), 041109 (Aug 12, 2013). ; http://dx.doi.org/10.1117/1.JEI.22.4.041109


Figures

Graphic Jump LocationF1 :

Overview of the video-based parking occupancy detection system.

Graphic Jump LocationF2 :

Overview of the offline (training) phase.

Graphic Jump LocationF3 :

Cross-validation test for determining histograms of oriented gradients (HOG) feature dimension for our vehicle classifier: (a) number of classification errors versus feature dimension and (b) number of support vectors in the resulting classifier versus feature dimension.

Graphic Jump LocationF4 :

Vehicle detection in the first frame using the classifier and two-dimensional sliding window search in the region of interest (ROI).

Graphic Jump LocationF5 :

Examples of unreliable frames for estimating parking occupancy illustrating (a) an occlusion and (b) a vehicle in the process of parking.

Graphic Jump LocationF6 :

Occlusion is detected when a foreground blob is partially inside the parking area of interest and partially outside the parking region.

Graphic Jump LocationF7 :

Candidate blobs detected due to changes produced by a parked vehicle (a–c), headlights (d–f), and a parked vehicle and its shadow (g–i).

Graphic Jump LocationF8 :

Parking angle detection is performed by finding the longest line in the detected vehicle: (a) current frame, (b) estimated background, (c) binary mask illustrating the location of the detected vehicle, and (d) longest line (green) identified in the image region corresponding to the detected vehicle.

Graphic Jump LocationF9 :

Fitting an ellipse to a blob. The minor and major axes of the ellipse are selected as the line segments that have the highest length ratio among all the perpendicular line segment pairs passing through the centroid.

Graphic Jump LocationF10 :

Stall parking. The parking space for each vehicle is separated by lines painted on the road surface, shown as dashed lines in the figure.

Graphic Jump LocationF11 :

Vehicle classification based on the area of detected blobs for parked vehicles: (a) current frame and (b) detected blobs for a long (truck) and short (car) vehicle.

Graphic Jump LocationF12 :

Vehicle classification by estimating the length of a vehicle.

Graphic Jump LocationF13 :

Exclusion zones (a) in front of a private driveway, (b) in front of a fire hydrant, and (c) on a road adjacent to an on-street parking area.

Graphic Jump LocationF14 :

Sample video frames illustrating the field of view of the cameras in the block (a) Camera 1, (b) Camera 2, and (c) Camera 3.

Graphic Jump LocationF15 :

Sample video frames illustrating different weather conditions at which test videos were captured.

Graphic Jump LocationF16 :

Sample video frames illustrating different parking scenarios.

Graphic Jump LocationF17 :

Parking boundary violation detection results. The vehicle inside the red box violates the law. The truck inside the white box is classified as a long vehicle and does not violate the law even though it passes over the boundaries of a parking stall.

Tables

Table Grahic Jump Location
Table 1Detection accuracy of the video-based parking occupancy detection method across 5 days of video captured from three different vantage points.

References

Shoup  D. C., The High Cost of Free Parking. , Vol. 7,  Planners Press, American Planning Association ,  Chicago  (2005).
Shoup  D. C., “Cruising for parking,” Transport Policy. 13, (6 ), 479 –486 (2006). 0967-070X CrossRef
Loce  R. P., Saber  E., “Fuel consumption in San Francisco,” http://spie.org/x1816.xml ( March 2013).
Wolff  J. et al., “Parking monitor system based on magnetic field sensors,” in  IEEE Intell. Transport. Syst. Conf. , pp. 1275 –1279,  IEEE ,  Toronto, Canada  (2006).
Mimbela  L. E. Y., Klein  L. A., “Summary of vehicle detection and surveillance technologies used in intelligent transportation systems,” http://www.fhwa.dot.gov/ohim/tvtw/vdstits.pdf ( August 2013).
Shaheen  S., Rodier  J. C., Eaken  M. A., “Smart parking management field test: a bay area rapid transit (BART) district parking demonstration,” Technical Report by Institute of Transportation Studies, UC Davis (2005).
Deruytter  M., Anckaert  K., “Video-based parking occupancy detection,” Proc. SPIE. 8663, , 86630O  (2013).CrossRef
Agarwal  S., Awan  A., Roth  D., “Learning to detect objects in images via a sparse, part-based representation,” IEEE Trans. Pattern Anal. Mach. Intell.. 26, (11 ), 1475 –1490 (2004). 0162-8828 CrossRef
Tsai  L.-W., Hsieh  J.-W., Fan  K.-C., “Vehicle detection using normalized color and edge map,” IEEE Trans. Image Process.. 16, (3 ), 850 –864 (2007). 1057-7149 CrossRef
Felzenszwalb  P. F. et al., “Object detection with discriminatively trained part-based models,” IEEE Trans. Pattern Anal. Mach. Intell.. 32, (9 ), 1627 –1645 (2010). 0162-8828 CrossRef
Dalal  N., Triggs  B., “Histograms of oriented gradients for human detection,” in  Proc. IEEE Computer Society Conf. Computer Vision and Pattern Recognition , Vol. 1, pp. 886 –893,  IEEE ,  San Diego, CA  (2005).
Felzenszwalb  P. F. et al., “Object detection with discriminatively trained part-based models,” IEEE Trans. Pattern Anal. Mach. Intell.. 32, (9 ), 1627 –1645 (2010). 0162-8828 CrossRef
Neubeck  A., Van Gool  L., “Efficient non-maximum suppression,” in  Int. Conf. Pattern Recognition , Vol. 3, pp. 850 –855,  IEEE ,  Hong Kong  (2006).
Stauffer  C., Grimson  W. E. L., “Adaptive background mixture models for real-time tracking,” in  Proc. IEEE Computer Society Conf. Computer Vision and Pattern Recognition , Vol. 2, pp. 252 ,  IEEE ,  Fort Collins, CO  (1999).
Oliver  N. M., Rosario  B., Pentland  A. P., “A bayesian computer vision system for modeling human interactions,” IEEE Trans. Pattern Anal. Mach. Intell.. 22, (8 ), 831 –843 (2000). 0162-8828 CrossRef
Lo  B., Velastin  S., “Automatic congestion detection system for underground platforms,” in  Proc. Int. Symp. Intelligent Multimedia, Video and Speech Processing , pp. 158 –161,  IEEE ,  Kowloon Shangri-La, Hong Kong  (2001).
Horn  B., Schunck  B., “Determining optical flow,” Artif. Intell.. 17, (1–3 ), 185 –203 (1981). 0004-3702 CrossRef
Lucas  B. et al., “An iterative image registration technique with an application to stereo vision,” in  Proc. Int. Joint Conf. Artificial Intelligence , Vol. 3, pp. 674 –679,  AAAI ,  Vancouver, Canada  (1981).
Huang  Y., Zhuang  X., “Motion-partitioned adaptive block matching for video compression,” in  Proc. IEEE Int. Conf. Image Processing , Vol. 1, pp. 554 –557,  IEEE ,  Washington, DC  (1995).
Makarov  A., Vesin  J.-M., Kunt  M., “Intrusion detection using extraction of moving edges,” in  Proc. 12th IAPR Int. Conf. Pattern Recognition , Vol. 1, pp. 804 –807,  IEEE ,  Jerusalem, Israel  (1994).
Paragios  N., Tziritas  C., “Detection and location of moving objects using deterministic relaxation algorithms,” in  Proc. 13th Int. Conf. Pattern Recognition , Vol. 1, pp. 201 –205,  IEEE ,  Vienna, Austria  (1996).
Cucchiara  R. et al., “Improving shadow suppression in moving object detection with HSV color information,” in  Proc. IEEE Intelligent Transportation Systems Conf. , pp. 334 –339,  IEEE ,  Oakland, CA  (2001).
NYC Department of Transportation, “Traffic violation codes in NYC,” http://www.nyc.gov/html/dof/html/parking/violation_codes.shtml ( January 2013).
Office of Documents and Administrative Issuance, Washington, DC, http://www.dcregs.dc.gov/Search/FullTextSearch.aspx?SearchType=DCMR&KeyValue=on+street+parking ( March 2013).
NYC Department of Transportation, “Parking tickets in NYC,” http://parkitnyc.com/parking-nyc/nyc-parking-tickets/ ( March 2013).
Illingworth  J., Kittler  J., “A survey of the Hough transform,” Comput. Vision Graphics Image Process.. 44, (1 ), 87 –116 (1988). 0734-189X CrossRef
Seber  G. A., Lee  A. J., Linear Regression Analysis. , Vol. 936,  John Willey & Sons ,  New York  (2012).
Chao  T.-H., Lau  B., Park  Y., “Vehicle detection and classification in shadowy traffic images using wavelets and neural networks,” Proc. SPIE. 2902, , 136 –147 (1997). 0277-786X CrossRef
Unzueta  L. et al., “Adaptive multicue background subtraction for robust vehicle counting and classification,” IEEE Trans. Intell. Transport. Syst.. 13, (2 ), 527 –540 (2012). 1524-9050 CrossRef
Morris  B., Trivedi  M., “Improved vehicle classification in long traffic video by cooperating tracker and classifier modules,” in  Proc. IEEE Int. Conf. Video and Signal Based Surveillance , p. 9 ,  IEEE  (2009).
Gupte  S. et al., “Detection and classification of vehicles,” IEEE Trans. Intell. Transport. Syst.. 3, (1 ), 37 –47 (2002). 1524-9050 CrossRef
Avery  R., Wang  Y., Scott Rutherford  G., “Length-based vehicle classification using images from uncalibrated video cameras,” in  Proc. IEEE Conf. Intelligent Transportation Systems , pp. 737 –742,  IEEE  (2004).
Zhang  Z., “A flexible new technique for camera calibration,” IEEE Trans. Pattern Anal. Mach. Intell.. 22, (11 ), 1330 –1334 (2000). 0162-8828 CrossRef
NYC Department of Transportation, “Exclusion zone parking tickets,” http://parkitnyc.com/parking-nyc/nyc-parking-tickets/ ( March 2013).
San Francisco Municipal Transportation Authority, “San Francisco parking occupancy detection performance requirement,” http://sfpark.org/wp-content/uploads/2011/09/SFpark_SensorPerformance_v01.pdf ( March 2013).

Some tools below are only available to our subscribers or users with an online account.

Related Content

Customize your page view by dragging & repositioning the boxes below.

Related Book Chapters

Topic Collections

PubMed Articles
Advertisement
  • Don't have an account?
  • Subscribe to the SPIE Digital Library
  • Create a FREE account to sign up for Digital Library content alerts and gain access to institutional subscriptions remotely.
Access This Article
Sign in or Create a personal account to Buy this article ($20 for members, $25 for non-members).
Access This Proceeding
Sign in or Create a personal account to Buy this article ($15 for members, $18 for non-members).
Access This Chapter

Access to SPIE eBooks is limited to subscribing institutions and is not available as part of a personal subscription. Print or electronic versions of individual SPIE books may be purchased via SPIE.org.