Research On Video Vehicle Detection And Tracking Method Based On Depth Model

With the continuous development of unmanned technology,human-computer interaction technology in video surveillance and traffic field,detection of video vehicle tracking target vision gradually occupies an important position in the field of intelligent video surveillance has become one of the key parties based on moving vehicle detection and vehicle access information type.Given the initial state of a target vehicle,the task of the target vehicle detection and tracking algorithm is to estimate the state of the target vehicle in the subsequent video.However,in the practical application of video vehicle detection and tracking,there are many interference factors that affect vehicle tracking efficiency,such as variability of target motion,vehicle scale,intensity of illumination,occlusion factor and complex background.Therefore,it is necessary to study a best robust and accurate video vehicle detection and tracking algorithm.The disadvantage of traditional video vehicle detection and tracking algorithm is the use of manually extracted underlying visual features to form the vehicle appearance model,which results in insufficient information in prediction models.According to the defects of traditional algorithm,by deep learning basic model has inspired great success in image classification and recognition task,the depth of network model is applied to the video vehicle detection and tracking algorithm,and through the study of the extraction of high deep features to construct a real-time appearance model.The specific research contents of this paper are as follows:First,combined with the research background and practical significance of video moving vehicle detection and tracking algorithm,this paper elaborates the research status of vehicle tracking algorithm at home and abroad,and puts forward a new idea of video vehicle detection and tracking technology.Second: according to the characteristics of the low efficiency of manual extraction vehicle detection and tracking process,as well as the attitude change,scale change,vehicle occlusion,effects of complex environmental adverse factors as the research subject,this paper introduces the idea of deep learning,deep learning theory,mainly discusses four kinds of common depth model,introduces its basic the thought principle and practical application of concrete,put forward a theoretical basis for the following algorithm.Third: in order to solve the problem of vehicle merging and shape distortion caused by video vehicle shadow points,a shadow detection algorithm combining Fast RCNN depth model is proposed.Firstly,in the video vehicle image,the Selective search method is used to extract the rectangular area of multiple vehicle candidates,and the Hessenberg decomposition method is used to separate the moving vehicle from its shadow area.Then we extract the shadow feature from deep network,detect the shadow with PCA analysis,train and optimize the network,and finally identify the vehicle area contained in the moving shadow,so as to achieve the effect of quickly removing the shadow.Fourth: due to the phenomenon of target drift and mismatch in target vision tracking process,a multi-target visual tracking algorithm based on stack noise cancellation self coder network and online Deep Boost learning(Online Deep Boost)strategy is proposed.This method first studies the local-global characteristics of the target through the ODB method on the SDAE network.Then the target is tracked and classified according to the particle filter and soft-max classifier based on the characteristic weight,and the most similar state value of the target is obtained.Finally,the time factor is introduced to calculate the dynamic duration of the target appearance,so as to update the appearance model to adapt to the change of the target appearance.