Traffic Congestion Classification

Traffic congestion is the most useful information for the drivers among all traffic information. To reveal the degree of traffic congestion, traffic flow, traffic density, and traffic speed are the important and useful traffic parameters. Thus, we process a traffic surveillance video to estimate the traffic parameters and classify the traffic congestion into five levels: jam, heavy, medium, mild, and low. In the following sections, methods of traffic parameters estimation and traffic congestion evaluation are proposed and described.

4.3.1 Traffic Parameter Estimation

In the proposed framework, three traffic parameters: traffic flow, traffic speed and traffic density are needed simultaneously to analyze the traffic congestion. Thus, we use the virtual detectors installed at initialization procedure to estimate the traffic parameters. On the basis of the bidirectional roadway analysis, the traffic parameters can be calculated for both directions of the roadway individually.

(1) Traffic Flow

Traffic flow Fl is defined as the number of moving vehicles passing through the scene in a time interval. In traditional methods, tracking all the moving vehicles on the roadway is a conventional way to calculate the flow. However, tracking all the vehicles on the roadway is extremely complicated and time-consuming for vehicle occlusion problem and lane changing behavior. Therefore, in our proposed framework, the virtual detectors on each lane of roadway are utilized for counting the traffic flow.

So traffic flow is defined as how many vehicles trigger the virtual detectors in a time

38

interval. The vehicle triggers the virtual detector when it passes through the detector.

To ensure that the vehicle truly triggers the detector, the foreground pixels of the vehicles have to occupy at least a portion (say a quarter) of the triggered virtual detector. This limitation can reduce the erroneous judgment caused by noises. The ratio can be changed with the quality of surveillance videos.

For the usage of virtual detectors, the vehicles are tracked when they trigger the virtual detectors, so the vehicle tracking on the whole roadway is unnecessary. A useful property is that only one virtual detector is installed on each lane, which simplifies the vehicles matching in vehicle tracking procedure. Because only one vehicle can occupy one virtual detector at the same time in normal situation, we just match the vehicle that is occupying the same virtual detector in two consecutive video frames for determining whether the two vehicles are the same. If they are the same vehicles, traffic flow Fl remains the same. Otherwise, it increases one. The color histograms of vehicles are used to match the vehicles here.

(2) Traffic Speed

Traffic speed Sp is the average speed of the moving vehicles in a time interval.

Generally speaking, speed is a ratio between moving distance and the time spent. To achieve this goal, it is necessary to track the vehicle for the length of its trajectories and to record the time to generate the trajectories. As discussed in previous sections, vehicle tracking on the whole roadway is not feasible in the complicated traffic situation. Hence, to obtain the speed of the moving vehicles, we estimate the speed of a vehicle when it triggers a virtual detector. In principle, a slower moving vehicle will trigger a virtual detector for more consecutive frames, but the situation for a fast moving vehicle is opposite. Hence, the speed approximation can be done by counting the number of frames that a moving vehicle triggers a virtual detector. Therefore,

39

traffic speed is refined as the average of frames for vehicles triggering the virtual detector in a time interval. The following equation is the definition of traffic speed S:

∑

where Fl is traffic flow, fv_i denotes the number of frames for that ith vehicle triggers the virtual detector, num_veh is the number of vehicles triggering the virtual detector in a time interval, and fps (frames per second) stands for sample rate of the surveillance video.

(3) Traffic Density

In general, traffic density is a ratio between the number of vehicles and the area of roadway. However, it is a difficult task to correctly segment all vehicles on roadway due to vehicle occlusion problem which is common at the far side of camera capturing.

Consequently, we choose another way to calculate the density of traffic.

After segmenting out all pixels of foreground image in current frame, the traffic density is the ratio between the number of pixels in foreground and the number of pixels of roadway. Thus, the more the foreground pixels are, the higher the traffic density is. Nevertheless, there is still a problem that the moving vehicle which is near the camera is much larger than it is far from the camera. This situation makes the defined density out of reality. Because a vehicle occupy the same ratio of width of roadway regardless of distance from camera, the ratio between foreground pixels and roadway are calculated row by row to reduce the influence of camera’s vision depth.

The average of the ratios for frames over a period of time is the traffic density Dens is defined below:

40

where num_fr is the number of frames in a time interval, g is the height of roadway mask, PFfi is the number of foreground pixels of row i in frame f, PRfi is the width of roadway mask of row i in frame f.

4.3.2 Traffic Congestion Evaluation and Classification

To evaluate the traffic congestion for traffic flow over a period of time, how to use the traffic parameters estimated during traffic monitoring is an important issue. In general, the higher the traffic density is, the more crowded the traffic is; the slower the traffic speed is, the more crowded the traffic flow is. Thus, the congestion evaluation can be designed on the basis of relation between the two traffic parameters. Therefore, the traffic congestion value Cong is defined as a ratio between traffic density and traffic speed. Thus, the higher the value is, the more crowded the traffic flow is. The congestion value becomes large while traffic density is larger and traffic speed is smaller. The equation of traffic congestion value is

Sp

Cong_day = Dens (25)

where Dens is traffic density and Sp is traffic speed, and the value will be used to make traffic congestion classification. As for traffic flow, it increases as the traffic congestion becomes crowded. However, it decreases when traffic is too crowded to move. Hence, the value is not directly employed to evaluate traffic congestion but used in traffic speed Sp evaluation.

41

For a clip of traffic surveillance video, congestion value is the only feature used for classification. Based on this feature, the traffic congestion levels are classified into five levels: jam, heavy, medium, mild, and low. In our framework, we adopt two methods for classification: Neighboring Class Distinguishing (NCD) and SVM classifier.

For NCD, the congestion values are divided into five intervals which stand for five levels of congestion respectively, and the congestion level of a congestion value is the corresponding interval where the congestion value locates. Moreover, thresholds, τd1, τd2, τd3, τd4, between five congestion levels are determined by training in advance.

As shown in Figure 15, for two neighboring congestion levels, the mean value of the average congestion values Cong_day in two level training data is the threshold between the two levels. For SVM classifier, the training data in five congestion levels are used to train a classifier model which is the base to classify the traffic congestion in testing video.

Figure 15. Determination of traffic congestion thresholds for NCD.

42

Chapter 5. Nighttime Traffic Congestion Classification

In this chapter, a nighttime traffic congestion classification framework is presented. The details of the module consisting of headlights detection, vehicle detection, and traffic congestion classification are described in the following sections.