Research On Identification And Tracking Technology Of Vehicle In Complex Scenes

In the current intelligent transportation system, recognition and tracking of vehicle is always thecore part. It can provide all kinds of dynamic information on transportation and environment whichcan contribute to unified management and dispatch. And it relieves congestion and reduces accidents.Therefore, accurate identification and long-term tracking of vehicle has been hot spots of study onintelligent traffic monitoring system. The paper focuses on the identification and tracking theory,discusses the method of identification and tracking of vehicle in detail in four steps, and uses specificexperiments to prove the reliability and effectiveness of the proposed algorithm. Specific work is asfollows:(1) The paper proposes an improved moving target detection algorithm based on Gaussian MixtureModel. The algorithm changes learning rules of traditional method by feedback of matching. Itovercome the drawback of fracture or separation when detection and excludes the interference ofvehicle and environment on background learning. Experiments indicate that the proposed method ismore accurate than the classic one on extracting moving target.(2) The paper proposes a shadow detection algorithm based on HSV color space and GMM. Themethod uses artificial sampling and HSV color space method to obtain shadow samples collectionwhich is used to estimate parameters of GMM by using EM(Expectation Maximization) approach.Finally the shadow model can be used to distinguish vehicle and shadow. The experiments indicatethat the method can distinguish vehicle and shadow effectively.(3) The paper uses seven Hu invariant distance、degree of dispersion、aspect ratio and compactnesswhich compose of10-dimensional shape feature vector and3-level BP neural network to classifypedestrians, carts, car, bicycle or motorcycle. The experiments indicate the trained neural networkclassifier can classify the four categories of target effectively.(4) The paper proposes an improved TLD tracking algorithm. It adds a Haar feature-based classifierbased on Adaboost method which is combined with the original classifier. They constitute asemi-supervised co-training classifier which improves the generalization capability of classification.The experiments show the method can improve the effectiveness of tracking.