The Research On Visual Multiple Object Tracking Algorithm Technology

As one of the most significant directions in the field of computer vision,visual multi-target tracking has been widely used in smart transportation,intelligent monitoring and other areas.However,due to the diversity of target scales,the number of tracking tasks and visual target identity error,this paper conducts research on target detection,tracking model establishment,trajectory management,data association and association clues,the main content of the paper is as follows:Firstly,the design of the multi-scale pedestrian detection algorithm is completed.After comparing the current commonly used detection algorithms,YOLOV3 is chose as the basic detection framework,and according to the principles of classification accuracy and network lightweightness,Res Net and SENet are selected.Global and local multi-scale information is embedded when constructing feature pyramids.The detection algorithm in this paper and three others are tested in different scenes,and the results verifies that our algorithm improve accuracy.Secondly,on the basis of object tracking model,a multi-target tracking algorithm based on Kalman prediction and detection association is proposed,generating the distance threshold by combining Kalman filter state and the detection result,and propose a method of constructing a loss matrix based on the spatial similarity measure,a lifecycle management mechanism is created to manage multi target.The experiments show proposed algorithm performs well,but ID switch needs to be discussed.Thirdly,in order to cope with the frequent change of pedestrian target identity(ID),apparent information as association clue is proposed.The paper focus on confirming IDS and designs re?id networks including object classification networks based on three common loss functions and re-identification networks based on discriminant and generative learning,the training on the large-scale reidentification data set has completed.Finally,based on Kalman prediction and detection of multi-target tracking,the apparent information obtained from the two designed networks is transferred to multi-target tracking,and the measurement of featue is introduced.A multi-layer data association mechanism based on spatial information and apparent information is proposed.Experiments indicate the integration of appearance features increases association clues,reduces the number of target ID jumps,and enhances the reliability of the tracking algorithm.Besides,the apparent features obtained from discriminative and generative learning are more reliable in association matching.