Traffic surveillance systems have been widely used for monitoring roadways in recent years. Unfortunately, the repository of the captured videos is so large that it is almost impossible to manually understand the contents of the videos. In fact, it is useful to utilize these traffic video data which can be processed to extract abundant traffic information for real-time intelligent transportation applications. Therefore, plenty of researches have been focused on automatic traffic events analysis such as traffic accidents, violation, and congestion. In this thesis, we investigate the event that the public concern most: roadway traffic congestion.
In the past, when traffic jams occurred, the police or drivers would inform the traffic control centers, and people detoured to avoid the traffic jams after radio station broadcasted that information. Nowadays, a variety of sensors such as loop detector, infrared detector, and Closed Circuit Television camera are used to gather the instant traffic information in traffic control system. However, the cameras are the particular devices that not only can observe the traffic situation but also record all events that happen on roadways all the time, which provides us with more plentiful traffic information. Moreover, due to the advantage of non-invasive installation, the cameras have distributed over all freeways and the main roadways in metropolises. Thus, it facilitates the possibility of the establishment of the complete traffic information. If a traffic surveillance system can automatically analyze the level of traffic congestion from traffic surveillance videos, the congestion message can be immediately provided for the public. Moreover, with the rapid development of intelligent mobile devices such as smart phone and personal digital assistant, drivers can earlier get traffic information and recommended alternate routes for avoiding traffic jam.
2
Aiming at the benefits arising from integration of video analysis technique and intelligent mobile devices, real-time classifying the traffic congestion in daytime and nighttime surveillance videos is the goal we are going to achieve in this thesis. The traffic congestion is classified into five levels: jam, heavy, medium, mild, and low. Jam is the situation that the vehicles fully occupy the roadways and almost all of the vehicles move slowly or completely stop. Heavy indicates that most of the vehicles on the roadway run slowly but seldom stop. In medium level, the difference from aforementioned levels is that all vehicles can move smoothly, and there still are many vehicles moving on the roadway. Mild means that the number of vehicles is much less than that in medium level and the vehicles move at normal speed. Low denotes that only few vehicles pass through roadway. Figure 1 shows the examples of video frames in five congestion levels from surveillance videos.
(a) (b) (c)
(d) (e)
Figure 1. Captured frames in five congestion levels. (a) Low. (b) Mild. (c) Medium. (d) Heavy in the right side of roadway. (e) Jam in the left side of roadway.
3
In order to analyze the traffic information from video, image processing techniques and knowledge of classification are essential. In general, procedure of analyzing video comprises selection of the region of interest, vehicle detection, vehicle tracking, and activity analysis. Nevertheless, in most of the existing works, a fundamental problem is that the performance of video analysis may not be stable with the varied environments. For example, a large number of vehicles may lead vehicle occlusion and cause the failures of vehicle segmentation. Moreover, vehicles which have similar features such as color, shape, texture, and moving direction increase the difficulty in vehicle tracking. On the other hand, a critical issue is that image processing is always time-consuming. For the requirement of real-time response, developing efficient frameworks and algorithms of video analysis is an important and inevitable challenge. Consequently, how to quickly and accurately evaluate the traffic congestion from traffic surveillance videos is the core problem in our work.
In this thesis, we propose a real-time traffic congestion classification framework which consists of daytime and nighttime modules to automatically process the daytime and nighttime surveillance videos for identifying the traffic congestion levels. For daytime surveillance videos, the moving vehicles on the roadway are extracted by background subtraction and shadow elimination technique. Afterward the extracted vehicles are used to calculate the important traffic parameters including traffic flow, traffic speed, and traffic density. Then the traffic parameters are utilized to evaluate and classify traffic congestion levels. For nighttime surveillance videos, the moving vehicles are detected by extracting and grouping the headlight candidates.
Subsequently a virtual detection line is utilized for evaluating the traffic congestion levels. Finally, we examine the proposed framework on real freeway surveillance videos captured at day and night data to demonstrate the accuracy and real-time response of traffic congestion classification.
4
The rest of this paper is organized as follow. Some related works on daytime and nighttime video processing are reviewed in Chapter 2. In Chapter 3, we introduce the proposed framework of traffic congestion classification. After that, we present the daytime module that includes initialization procedure, vehicle detection and traffic congestion classification in Chapter 4. In Chapter 5, the module of nighttime traffic congestion classification which is composed of headlight extraction, vehicle detection and traffic congestion classification is described. The experimental results of daytime module and nighttime module are shown and discussed in Chapter 6. In Chapter 7, we conclude this thesis and discuss the future works.
5