Research On Multi-target Tracking And Feature Extraction Of Vehicles In Traffic Video

With the development of artificial intelligence,intelligent traffic video system is intended to apply computer vision techniques to process the moving object detection automatically.However,the research of moving target detection and vehicle track:ing is still facing many challenging issues because of complex background,a lot of noise,irregular vehicle moving much occlusion of many objects,interference of analogues and background.Facing much more traffic video information,how to analyze vehicle feature and retrieve the information we need is becoming a hot issue.This paper studies the moving object detection,the multiple vehicle tracking and object feature extraction.About moving vehicle detection,there were some problems existing in the Vibe based algorithm in traffic video processing such as ghosting shadow appeared in video processing,difficulty of removing residual shadow of slow moving targets,and poor detection accuracy and robustness.This paper proposes an improved Vibe algorithm.It uses the gray-scale spatial information to build matrix of pixel life length to make ghosting or target’s residual shadow quickly blended into the samples of the background.It also combines with the OTSU method to set the dynamic threshold.The improved algorithm takes the good post-processing method to restrain dynamic noise.The paper also proposes several quality evaluation criteria based on statistics index of classification algorithms.Experiment results showed that the improved algorithm removed the ghost shadows and restrained the residual shadow of moving objects within less frames.It also promoted the accuracy and overall performance of moving object detection.Effective vehicle detection also provides a good basis for multiple target tracking and vehicle feature extraction.About multiple vehicle tracing,in order to solve the problem of object occlusion,sticking or separation,interference of similar objects and irregularity of object motion a region matching and tracking algorithm based on cascaded classification and SVM secondary recognition is proposed in this paper.According to the HOG-based cascade classification algorithm,the vehicle tracking region is effectively identified,it can reduce the impact of vehicle connectivity domain adhesion.The SVM classification algorithm based on LBP feature is added to remove the interference target and analogues,getting the highest confidence of the vehicle tracking box.According to the region assoceiation algorithm to track the tracking box stably.Exqperiments show that the multiple target tracking algorithm can track the vehicle continuously and stably,and has high accuracy.About object features and retrieving the information,this paper designs a traffic video retrieval comparison framework based on vehicle features.Firstly,the features of multiple vehicle tracking are analyzed,According to the HSV non-uniform quantification principle,the main area color of the target vehicle is extracted,The Bayesian classifier is used to classify the vehicle features.And the features of the vehicle will be structured storage.A new vehicle search and matching model with color and vehicle feature is proposed,according to the inverted index to retrieve the comparison,quickly locate the similar vehicles we need to find.The validity and accuracy of feature extraction and retrieval are proved by experiment results.