Research On Vehicle Abnormal Behavior Detection Methods In Urban Traffic Scene

With the rapid development of China's national economy, the living standards of the people have been continuously improved. Automobile are already not a luxury that can not be imagined and it has entered into thousands of households. The car ownership of China has reached to 172 million by the end of 2015. The sharp increase in car ownership has brought great convenience to people's daily travel, but also has brought a serious challenge to traffic management in city, such as urban traffic jam and frequent traffic accidents, which has caused a large number of casualties and property losses to people. Vehicle abnormal behaviors refer to vehicle violation behaviors, which means automobile episodic events on the road, including running a red light, pressing line, illegal steering etc. According to statistics, there are more than one billion people died in all kinds of traffic accidents in China over the past ten years, which is far higher than the western developed countries, and more than 90 percent of the accidents are caused by abnormal behavior of the vehicle. Therefore, the detection and early warning of vehicle abnormal behaviors has become a hot research problem of the traffic management department. The main content of this paper is to study the detection methods of vehicle abnormal behaviors in urban traffic scenes, specific as follows:First of all, the traffic signal lamp state detection method based on color space model is studied. Color histogram feature and Bhattacharyya coefficient of signal lamp were used to locate signal lamp area. Then, color space model and color space threshold method were used to detect the signal light in traffic scene and we compared the status recognition rate of the traffic signal lights in the RGB color space, HSV color space and YCrCb color space. The experimental results show that the detection method of the signal lamp baesd on HSV color space is better than RGB color space and YCrCb color space.Secondly, geometry and texture features in the data vehicle image is analysed and vehicle area detection method based on the contour of the vehicle and the symmetry of the vehicle license plate is studied in urban traffic scenes. Also, some other vehicle detection methods such as method based on the characteristics of licence plate, method based on Gabor feature and method based on Haar-like feature are discussed as a comparison at the same. Experimental results show that the method based on the contour of the vehicle and the symmetry of the vehicle license plate in urban traffic scenes is better than method based on the characteristics of licence plate, method based on Gabor feature and method based on Haar-like feature.Finally, a vehicle behavior detection method based on combined feature in urban traffic scenes was proposed, which HOG feature combines with LBP feature in series, and the support vector machine (SVM) is used for automatic classification. Some other feature extraction methods such as HOG (Histogram of Oriented Gradient) feature, LBP (Local Binary Pattern) feature and EOH (Edge Orientation Histogram) feature were proposed as a comparison. Experimental results show that combined feature based on HOG feature and LBP feature has a higher recognition rate, up to 93.6%. Red light running has the highest rate of decision, and the illegal behavior is the most difficult to determine.