Data Association Multi-target Tracking Based On Deep Learning

Nowadays,with the great development of computer,the technology of video information analysis has officially entered the era of intelligence,the demand for video target tracking is growing day by day.Multi-target tracking technology plays an important role in the field of computer vision and is widely used in security,military,autonomous driving,traffic monitoring and other fields.Multiple Object Tracking(MOT),the main task is to detect and identify a specific target in a given image sequence and to calibrate the motion trajectory of each target.However,in the actual environment,the change of lighting,camera motion,the change of the target's own characteristics,the target occlusion,the similarity between the target and the background color,and the chaotic background all increase the difficulty of designing the multi-target tracking algorithm.Over the past five years,tracking by detection's multi-target tracking algorithm has gone mainstream as its effectiveness has improved.The high-precision deep learning detector is used to detect and identify the image,and the detection result is used as the input,and the targets between adjacent frames are matched by the method of data association,so as to obtain the target track.This kind of method can reduce the difficulties caused by complex tracking background and moving target deformation,but it is difficult to deal with the problems such as long-term occlusion and number exchange when the target overlaps due to the correlation of only adjacent frames.This topic focuses on the target detection technology based on deep learning,test data correlation multi-target tracking method and target motion prediction and block processing are studied,proposed a data correlation multiple target tracking system based on deep learning framework,In order to significantly enhance the accuracy of multi-target tracking,reduces the residual and checked by mistake.The main work includes the following four parts:1.According to the frontier development of target detection based on deep learning,the high-precision multi-target detection algorithm is compared and studied to provide reliable input for the multi-target tracking system.2.Summarized the commonly used methods of data association,including the method of association between adjacent frames and the method of association between multiple frames.3.The deep learning-based RPN network is studied,the detector of deep learning is used to predict the target position,and then the multi-objective data association algorithm is carried out by KM algorithm,which improves the accuracy of data association.4.The occlusion processing model is studied when the target is blocked for a short time.The kalman filter is used as the motion model to predict the position of the next frame of the blocked target.The main innovations of this dissertation are as follows :(1)the deep learning detector is used as the tracker to predict the target position,and then the KM algorithm is used for data correlation,which effectively utilizes the high-precision detector to extract the appearance characteristics of the target and improves the tracking accuracy.(2)a model to deal with target occlusion is proposed,which combines the features extracted from twin networks with the predicted position of kalman filter,and uses KM algorithm matching to deal with the problems of track interruption and target omission.