Research On Multi-target Tracking Algorithm Based On Deep Learning

Moving target detection and video tracking technology have a wide range of application scenarios in many fields of civil and military use.At present,although the use of cameras in public places such as stations,banks,and shopping malls is very common,the actual monitoring tasks for pedestrian tracking still need massive human intervention due to complex background and crowd environments.Because of its academic and commercial potential,multi-target tracking has gradually attracted attention in computer vision.The main purpose of studying multi-target tracking is to reduce the dependence on humans,and to automatically complete the real-time observation of targets in complex environments and the analysis and description of the behavior of interested targets.In recent years,the development of deep learning and other technologies has greatly improved the multi-target tracking algorithm,but when the target is deformed,scaled,occluded,etc.,it is still easy to cause tracking failure.The existing real-time Multi-Object Tracking(MOT)research mainly focuses on correlation determination,which is essentially a real-time association method between detection and tracking objects,not an entire MOT.In order to improve the speed of the algorithm,this paper integrates multiple modules of multi-target tracking into a neural network.This fusion has two advantages: on the one hand,it reduces redundant calculation by sharing the underlying features of the image;on the other hand,it reduces the time complexity of the algorithm by starting from the feature map of the image rather than from the pixel level.Based on deep learning technology,this paper proposes a new fusion network,and proposes an improved scheme for trajectory prediction module and data association algorithm in multi-target tracking.The main contents of this paper are as follows:(1)A multi-target pedestrian tracking algorithm based on detection,re recognition and social LSTM is proposed.Aiming at the influence of visual angle,illumination,noise and other factors on multi-target tracking,the network structure is designed to improve the generalization performance of the model;the pedestrian re recognition function is realized on the basis of the detection model;for the motion trajectory prediction module,the social LSTM is used to capture the interaction between targets to improve the prediction ability of the algorithm.The experimental results show that the algorithm proposed in this paper has significantly improved the running speed and tracking accuracy.(2)A multi-target pedestrian tracking algorithm based on multi module fusion is proposed.In order to further improve the real-time performance of mot algorithm,the functions of detection,pedestrian re recognition and trajectory prediction are integrated into a deep neural network,and the data association algorithm is improved.At present,the mainstream multi-target tracking algorithms need to use Kalman filter in the prediction process or Hungarian algorithm matching in data association,which need to transfer the output of neural network to CPU.The interaction between GPU and CPU will consume a lot of time and increase the running time of mot algorithm.The robustness of the algorithm is improved by introducing target features into the trajectory prediction module.Experimental results show that compared with the existing algorithms,the proposed algorithm has significantly improved the time complexity and the false matching rate.