Chapter 6. Experimental Results and Discussions
6.1.5 Performance of Execution Time
As shown in Table 11, the time for estimating traffic parameters is around 0.1 second per frame regardless of different congestion levels. In particular, the time for processing a frame is increased just a little as the traffic flow becomes crowded. That means the usage of virtual detectors is successful in reducing the computation cost as
65
the number of moving vehicles on roadway increases. As a whole, the proposed framework works well for the traffic surveillance video that processes at least 4 frames per second. So, using the key frames of videos is enough to accomplish the traffic congestion classification.
Table 11. Time for processing one frame in five congestion levels in daytime video.
Traffic
congestion Low Mild Medium Heavy Jam
Time (s) 0.08 0.09 0.10 0.10 0.11
6.2 Nighttime Traffic Congestion Classification
In this section, we are going to show the experimental results of the proposed nighttime traffic surveillance congestion classification module. The experimental results of headlight detection, vehicle detection, and traffic congestion classification are shown in the following sections. Moreover, the nighttime traffic surveillance videos that are captured on freeway by Taiwan Area National Freeway Bureau [44]
are used as the experimental data. The resolution of video frames is 352×240, and the sample varies from video to video. For the experiment, we use a computer with AMD 2.8 GHz dual-core CPU and 2.0 GB memory.
6.2.1 Headlight Detection
In this section, we present and discuss the results of headlight detection. The setting of the threshold: thY for luminance is 220 and thCV for color variation is 20.
These parameters are mentioned in Section 5.1. Figure 25(a) shows the original scene
66
of nighttime traffic surveillance and a horizontal green line denotes the virtual detection line, Figure 25(b) presents the detected bright region. Figure 25(c) illustrates the mask of detected bright blobs that touch the virtual detection line. The cyan bright blobs are the blobs which are filtered after headlight shape validation, but the red blobs are not.
Due to the use of virtual detection line, the headlight detection needs not to be applied to the whole bright pixels in the frame. Thus, only the bright pixels near the detection line are processed. Through the results, it is obvious that all headlights are are extracted after bright region detection. Some bright blobs not belonging to headlights, such as reflection of light on the roadside fence in Figure 25(5) and neon lamps of a bus in Figure 25(7), are also extracted. Then, shape validation is executed to discard these bright blobs and reserve other blobs as the headlight candidates for further vehicle detection.
67
1
2
3
4
5
6
7
(a) (b) (c)
Figure 25. Experimental results of headlight detection. (a) Original frame captured at night. (b) Detected bright regions. (c) Bright regions after shape validation.
68
6.2.2 Vehicle Detection
In this section, we demonstrate some results of vehicle detection at night. The setting of the threshold: Wexpected for expected vehicle width is 20 and Eallowable for error tolerance is 10 as mentioned in Section 5.2. Figure 26(a) shows the original scene of nighttime traffic surveillance and a horizontal green line as the virtual detection line. In Figure 26(b), the headlight candidates surrounded by a yellow rectangle are the actual headlights of vehicles and some solitary headlight candidates are removed. The horizontal line between two headlights indicates that the pair of headlights is a single vehicle. Through the results in Figure 26, we can find out that some headlight candidates that do not belong to the vehicle are removed by the vehicle detection as shown in Figure 26(2), Figure 26(3) and Figure 26(6).
1
2
3
69
4
5
6
7
8
(a) (b)
Figure 26. Experimental results of vehicle detection by headlight grouping. (a) Original frame captured at night. (b) Detected vehicles at virtual detection line.
70
6.2.3 Traffic Congestion Classification
We are going to present and discuss the experimental results of classification for nighttime traffic congestion levels in this section. Our approach is used to process the moving vehicles captured from vehicle front for headlight detection. Thus, only the incoming traffic flows can be analyzed in the nighttime surveillance videos. For the outgoing traffic flows, we can adopt another camera with opposite shooting direction to capture the traffic flows from vehicle front. In the experiment, there are 165 video clips, which monitor the unidirectional incoming traffic flows, are captured on the freeways from surveillance cameras during night. We use 49 video clips to train the thresholds for different congestion levels, and others are used as the testing video. The length of each video clip is 60 seconds.
The distributions of training data and testing data are shown in Table 12. The ground truths of the experimental data are determined manually according to the principles in Chapter 1. The traffic congestion Congnight of each training video clip in five congestion levels is estimated by our proposed framework, and all the results are used to determine the thresholds between the different levels by using the approaches mentioned in Section 5.3.
Table 12. Distribution of experimental nighttime surveillance video data.
Congestion Levels Low Mild Medium Heavy Jam
No. of Testing data 21 21 47 21 6
No. of Training data 9 9 20 9 2
71
(1) Neighboring Class Distinguishing (NCD)
Based on the traffic congestion Congnight estimated by using virtual detecting line, accuracy of traffic congestion classification is 95.4%. The accuracy is calculated by the Eq.(41). The confusion matrix of the classification is shown in Table 13. From the results, we discover that traffic congestion Congnight may be mis-classified into the neighboring traffic congestion levels. This means that Congnight is a reliable value to express the congestion situation because of no absurd mistakes. For instance, the low traffic flow is mis-classified into jam level.
Table 13. Nighttime traffic congestion classification results by Congnight (NCD).
Results of classification
Low Mild Medium Heavy Jam Accuracy
Ground Truth
(134)
Low(21) 21 0 0 0 0 1.00
Mild(21) 0 18 3 0 0 0.86
Medium(47) 0 3 43 1 0 0.91
Heavy(21) 0 0 0 21 0 1.00
Jam(6) 0 0 0 0 6 1.00
(2) Support Vector Machine (SVM)
The accuracy of congestion classification by using SVM classifier is 88.2%. The result is lower than that by using NCD due to the different thresholds determination methods. For the properties of that the traffic congestion degree is linear (i.e. from low to jam), each threshold between two levels is decided by only the two level training data is better than by the training data in all the five levels.
72
Table 14. Nighttime traffic congestion classification results by Congnight (SVM).
Results of classification
Low Mild Medium Heavy Jam Accuracy
Ground Truth
(134)
Low(21) 20 1 0 0 0 0.95
Mild(21) 0 11 10 0 0 0.52
Medium(47) 0 1 44 2 0 0.94
Heavy(21) 0 0 0 21 0 1.00
Jam(6) 0 0 0 0 6 1.00
6.2.4 Performance of Execution Time
The use of virtual detection line is successful in decreasing the time for processing frames in all congestion levels. In the experiments, the time for estimating traffic parameters is around 0.1 second per frame regardless of congestion levels. As a whole, the proposed framework works well for the traffic surveillance video that processes at least 4 frames per second. The result obviously shows the real-time response of our proposed framework. In addition, the size of headlight becomes a key factor to influences the execution time. It takes more time to process the larger headlights. As shown in Table 15, the jam level spends less time than that in the medium level due to the headlight size. That is the size of headlights has larger effect than congestion degree for the nighttime performance.
Table 15. Time for processing one frame in five congestion levels in nighttime video.
Traffic
congestion Low Mild Medium Heavy Jam
Time (s) 0.096 0.104 0.105 0.095 0.102
73