Research On Multi-source Video Sequence Target Tracking Algorithm Based On Machine Learning

In recent years,with the rapid development of intelligent video analysis and the country's emphasis on public security surveillance,people are deploying a large number of cameras to cover more areas and expand the scope of surveillance.Therefore,it is essential to realize intelligent identification and tracking of targets in multiple video sources,and has broad application prospects and practical value.In a real-world surveillance environment,the range of sight between multiple cameras usually does not overlap,and it is difficult for the cameras used for surveillance to capture clear faces.Only the overall appearance characteristics of pedestrians can be used to find and determine the person being tracked.However,the appearance is susceptible to various environmental factors,such as lighting,viewing angle,clothing,body posture and occlusion issues.The appearance characteristics of the same person in video images taken by different cameras will change significantly.The appearance characteristics of different pedestrians may be more similar than the same person,which will have a certain negative impact on target recognition and tracking.Therefore,aiming at the above-mentioned difficult problems,this paper focuses on the problem of automatic recognition and continuous tracking of the same pedestrian target in video sequences shot by multiple cameras in non-overlapping views.The main innovations of this paper are as follows:(1)In order to improve the accuracy of target tracking,we proposed an adaptive fusion target tracking based on joint detection algorithm.The deep and shallow convolutional features acted on the correlation filters separately to obtain response scores according to their respective advantages.The response scores of different convolutional features were adaptively fused by minimizing the loss.Then it combined with the location detection method of this study to judge the validity and authenticity of the predicted location,so as to get the optimal target tracking results.Finally,drew the corresponding motion trajectory for the tracked target.(2)In order to improve the accuracy of person re-identification,a multi-source video sequence person re-identification algorithm based on temporal attention was proposed in this paper.Firstly,the ResNet-50 network was used to extract the features of the input video sequence frame by frame.A series of frame-level features were input into the temporal attention network to generate corresponding weight scores.Then the frame-level features were weighted and averaged to obtain the sequence-level features.At the same time,to avoid weight scores from being aggregated in one frame,frame-level regularization was used to limit the difference between frames.Finally,the optimal results were obtained by minimizing the losses.The experimental results show that the method proposed in this paper can accurately re-identify and continuously track the target person as a whole,also can clearly reflect the moving direction,trajectory of the target and its accurate position in each camera view in the monitoring area.It also had excellent performance for angle of viewing variation,human posture variation,interference of similar targets,occlusion and scale variation.