Research On Vehicle Recognition In Intelligent Transportation

Intelligent Transportation System is an key part of modern transportation, theautomatic vehicle recognition is an important research direction in the filed ofIntelligent Transportation System. It is mainly used in highway fees, parkingmanagement, road traffic regulation and many other fields. For vehicle recognitionbased on image processing technology and pattern recognition with easy installation,wide detection range, extracting information rich etc., is a research hot spot and developtrend. A new vehicle recognition system based on video image sequences has designedin this thesis.This thesis mainly divided into three parts, which include moving vehicle detectionin the video sequence part, feature extraction part and vehicle recognition part.In the moving vehicle detection part, this thesis puts forward a new algorithm ofbackground reconstruction and update base on the analysis of the existing variety ofvehicle detection algorithm. This algorithm combines three frame difference methodand statistic method, it can quickly obtain a clean background, and can eliminate theinterference brought by the traffic volume excessive; while the background model canbe adapted quickly to the changes of scene. Background subtraction method can get themovement area of the vehicle, using the normalized cross-correlation function toeliminate shadow of vehicle in the gray space, this could get a good segmentationresults in practical application.In the feature extraction part, this thesis selected the SIFT feature to represent theimage after comparing multiple image feature, SIFT is robust for object scale, viewingangle and light changes. I abandon the Lowe’s keypoint detection method which istime-consuming, using evenly grid sampling to obtain the keypoint, and then generateSIFT feature descriptor in the image block of specified size. Not only extract featuresfaster, also taking into account the global feature of the image, and the experiment resulthas proved that evenly grid sampling SIFT feature performance better than commonSIFT feature.In the vehicle type part, I combine bag-of-word model and Support Vector Machine together to divided the vehicle into three types，which include cars, vans andtrucks. Represent the image into a bag of feature, then send the feature vector into SVMto learning and testing, we can finally obtained the type of vehicle. The experimentalanalysis of the influence of different parameter settings on the classificationperformance, and has shown that in the case of parameters to achieve the optimal, theproposed method can achieve a recognition rata of92%. Finally the thesis designed asimulate platform of vehicle recognition system.