Identification Method Of Vehicle Behavior Based On Surveillance Video

Recent years,with the development of computer and image processing technology,surveillance video is more and more widely used in traffic management.Vehicles' behavior,such as sudden lane changing and pushing into lane,would cause traffic accident and traffic jam,which were the significant cause of transportation safety and traffic order.So the method of detecting vehicles' lane changing behavior and pushing into lane would become a prop for traffic safety forewarning and law enforcement,aiming at lower the number of traffic accident and relieve traffic congestion.The main work is as follows:Firstly,vehicle detecting algorithms via HOG feature and SVM classification was used in this thesis.More than ten thousands of positive images and over thirty thousands of negative images were collected.After HOG feature was extracted SVM classifier would detect target cars in video.Secondly,an Improved Extended Kalman filtering algorithm was set to track vehicles.Otherwise,a district limitation was defined on the condition,optimizing vehicle trajectory and improving the accuracy of vehicle tracking.In this thesis,the DLT(Direct Linear Transformation)calibration method was used to calibrate the actual road scenery to get physical coordinates of the vehicle.Moreover,vehicle motion feature extraction method was built.After vehicle trajectory sequences were got,according to the fundamental principle,after preprocessing,improved K-means clustering and polynomial fitting process,the vehicle motion sequences were obtained.At last,through vehicle motion model vehicle motion feature sequences were finally earned.Lastly,Hidden Markov Model was fund to analyze vehicle motions.According to vehicle direction angle features,forward and backward algorithm connected with Baum-Welch algorithm,the next time state of vehicle was only determined by the right time state of vehicle.This thesis also utilizes SVM to predict vehicle motion.Real experiments demonstrate that the accuracy of HMM prediction on real road experiment could meet the accuracy need.The research achievements would lay a theory foundation and also practice basis for vehicle motion behavior analysis based on surveillance video.