Research On The Fast Vehicles Retrieval Algorithm In Surveillance Video

The development of Intelligent Transportation Systems (ITS) offers a new way tosolve urban traffic problems, it is a great potential for development of trafficmanagement and control techniques. ITS can analyze the traffic flow, to provide datafor urban road planning, at the same time can save the time of dealing with trafficaccident cases. So ITS has broad application prospects in road traffic monitoring,traffic accident scene investigation, public security and other aspects of theinvestigation.Intelligent Transportation Systems need to extract the interested informationfrom vast amounts of traffic surveillance video, so content-based video retrieval is anintelligent transportation system core technology. Content-based video retrievalsystem extracts the feature from the image, lookup in the video database, to retrievethe video segments with similar characteristics. In general, the video classificationsystem makes use of image descriptor based machine learning to classify differentobjects and scene. The application for content-based video retrieval in intelligenttransportation systems is retrieving a specific vehicle in traffic surveillance video, andcompared to conventional video retrieval system has its unique characteristics.This paper gives a detailed analysis of vehicle retrieval process, and studied therelated algorithms and implementation technology, and provide some method formotion region detection, feature extraction and classification of vehicles.In this paper, the main research work is as follows:(1) About moving vehicle detection, using the spatial correlation and temporalconsistency of the pixel in video sequences, this paper presents a compressed domainmoving object extraction algorithm fusion of Gaussian background modeling andgraph cut algorithm. This method combines the Gaussian background modeling andgraph cut algorithm respective advantages, achieved regional initialization byGaussian mixture background modeling, obtain initial movement area; finally usinggraph cut algorithm to correction the initial target area，remove the noise. Alloperations are conducted in the compressed domain, ensures that the method can fastand robust to extract the moving vehicle from h.264video coding standard codestream.(2) For vehicles feature extraction and classification problems, from theperspective of image feature points matching, will randomly Ferns algorithm isintroduced into vehicle video retrieval system, combined with Bayesian classifier, aalgorithm on multi-view video vehicle fast retrieval integration of multiple Bayesiandecision-making is proposed. Based on the Random Ferns algorithm, by extracting thetarget vehicle’s multiple image features for training, two bayes classifiers have beenbuilt. Then, the joint probability of these feature points which are detected in thevideo image is computed, and assign class lable to the video image to complete the mult-view video retrieval.