Vehicle Abnormal Behavior Analysis Based On Image Processing

With the rapid development of Chinese economy,there are more and more vehicles on the road.While enjoying the convenience of vehicles,we are also threatened by the frequent occurrences of traffic accidents.The problem of traffic safety has risen to the human level.However,most traffic accidents are caused by drivers' abnormal behaviors.The current video monitoring system is mainly used for obtaining evidence after the traffic accident happened,and it relies heavily on human's search,and thus leading to traffic accidents can only be treated as a postprocess rather than as an early precaution.To regulate vehicle driving behavior and reduce traffic accidents,the detection of traffic abnormal behavior in video monitoring has gradually become a research hotspot in the field of intelligent transportation.This paper takes the analysis of traffic abnormal behavior as the main direction.To discuss the behavior of driver's not wearing seat belt,using fake license plate and vehicle trajectory violation in traffic monitoring system,we also make solutions for the detection of three types of abnormities.The main contents are as follows:(1)For the detection of the safety belt,we propose a method of safety belts detection algorithm based on deep learning,it can quickly locate and detect safty belts through a dual cascade network.In order to realize the positioning from the car window to the main driving area step by step,two detectors are trained for the window area and the main driving area respectively by Faster RCNN.At the same time,we optimized the network structure of AlexNet.Then,a binary classifier is trained by using the improved AlexNet deep convolutional network to detect the safty belt in the main driving area.According to the experimental data,the algorithm of safety belt detection which is used in this chapter has the better effect,and the robustness of the system is better.(2)For the detection of using fake license plate,we design a system to realize license plate location,recognition and fake license plate detection.We also propose three criteria for the detection of fake-licensed car.First,the weight updating of traditional Adaboost algorithm are optimized,which can effectively prevent the over-fitting of the classification model.Then extract the HOG feature of the target region to train the improved Adaboost classifier,to realize the positioning from the car face to the license plate.After preprocessing of the license plate,the precise segmentation of characters is realized by using the eight connected domain and the vertical projection.Then the license plate recognition is accomplished by using the improved Adaboost classifier with the outline and structure feature of the characters.Finally,the detection of fake-licensed car is realized through three criteria: the information comparison of the license plate,the color comparison of the car and the spatio-temporal relationship based on the grid monitoring.(3)For the detection of the vehicle's trajectory violation,we use the KCF algorithm to achieve the fast tracking of targets.Then,we propose the criteria which are based on the analysis of vehicle trajectory,for judging the three types of violations:vehicle retrograde,illegal lane change and illegal turning.In the stage of vehicle tracking,the matrix operation is transformed into the dot product operation of DFT by using the cyclic matrix,which greatly reduces the computation and improves the efficiency of target tracking.Then,the kernel function is used to map the low dimensional nonlinear function into a high dimensional linear function,which makes the target feature linearly separable.Through the experiment we compared the accuracy and tracking error of four tracking algorithms.After obtaining the vehicle's track points,the trajectory coordinates are analyzed.And we use Hough transform to fit the lane line and the vehicle trajectory line.Finally,through analyzing the correlation between coordinates and lane lines,the detection and determination of three types of vehicle violations are realized.