Research On Vehicle Recognition And Adhered Vehicles Segmentation Methods In Traffic Video Surveillance

Intelligent Transportation System (ITS) is the key issue of the current worldtransportation field. It integrates advanced information technology, data communicationtechnology, electronic sensing technology, computer technology and all kinds of otheradvanced technologies synthetically and applies them to the ground traffic managementsystem, and has become an important development aspect of the modernized transportationsystem in21st century. While getting accurate real-time traffic parameters by videosurveillance method is a prerequisite for intelligent traffic management,the importance of intelligent traffic monitoring system based on video images, whose mainresearch target is to detect, recognize and track vehicles and analyze there behaviors so thatwe can get the important road information to make the traffic management intelligent, havebecome increasingly prominent.This paper mainly the approach of target vehicle recognition and the method how to splitadhesion vehicles, and the main works are as follows:Firstly, Base on the previous research, the structure of Haar-like rectangular feature andthe integral image which can compute the value of the feature is analyzed.Secondly, we studied a road vehicle method based on Haar-like feature. We useHaar-like features to describe vehicle feature and construct weak classifiers，and AdaBoostalgorithm is chosen to train the vehicle classifiers. Focusing on analysis of the trainingprocess of Adaboost algorithm, and against the time-consuming in the iterative processing,we proposed a sample’s Mahalanobis distance-based classifier selection methods. Meanwhile,compute the chosen classifier’s threshold and error rate based on the probability distributionof the positive and negative samples. And then update sample’s weights among the partition.The propose algorithm greatly reduces the complexity of the training process.Finally, this paper proposed a solution to the problem of vehicles adhesion: a regionsplitting method based on concavity analysis. First, detecting the vehicle region of interest(ROI), and analyzing the presence of adhered vehicles according to the convex hull and theblob analysis. For the adhered vehicles, we seek the splitting points (concavities pixels) usingscanning method, and propose a series of the rules to determine the optimal split line. The proposed method can split the adhered vehicle effectively.