Research On Lightweight Multi-target Tracking Algorithm Based On Joint Detection

With the widespread application of deep learning in the field of object tracking,the accuracy of multi-object tracking algorithms continues to improve,and they have become more robust and reliable.However,there are still many problems in the practical application of multitarget tracking technology in crowded scenes,such as occlusion problem,sometimes need to track the occluded part of the target;Association problem,in multiple consecutive video frames,need to associate the same object;There are also similar target jamming and target rerecognition problems.In addition,the multi-target tracking model based on deep learning generally has good tracking effect,but it is difficult to deploy them on the hardware platform with limited computing power due to its large amount of computation and redundant network parameters.In order to solve the above problems,this paper takes Fair MOT algorithm as the basic framework to expand and improve,and makes overall lightweight of the improved model to adapt to the development trend of intelligent devices in the future.The main work is as follows:Aiming at the problem of the number of identity switching caused by occlusion in dense environment,this paper makes an in-depth analysis on the multi-object tracking algorithm Fair MOT,and then designs a pedestrian tracking algorithm based on attention mechanism and data association.By introducing the attention mechanism into the high resolution feature extraction network,the algorithm extracts the high resolution features with more expressive power and strengthens the training of the re-recognition branch.Then,the data association module of pedestrian multi-target tracking is studied.As the pedestrians in the video walk continuously,there will be pedestrians leaving the picture,which requires matching and correlation of targets in different frames.In order to better match the target between frames,BYTE association strategy is adopted in the tracking module.Finally,the ablation experiment and comparison experiment prove that the proposed method can effectively solve the problem of pedestrian occlusion,reduce the number of missed detection,and improve the accuracy of the algorithm.Aiming at the problem of large number of network parameters and large amount of computation,this paper designs a multi-target tracking lightweight network model with high resolution features on the basis of the above.This model has the advantages of simple structure,fewer parameters and high resolution.It has faster reasoning speed than large networks.Through the analysis of the network model with added attention mechanism,it is found that it increases the complexity of the network and makes the channel between the layers of the model have a certain redundancy.Firstly,the human posture network model is studied in depth.The deep separable convolution method and lightweight network Shufflenet V2 are used to improve the network with added attention mechanism.In this way,the lightweight network can be guaranteed without losing its detection accuracy,so as to achieve the balance between lightweight and high expressiveness.Finally,the lightweight network is trained and tested on the multi-target tracking official data set.The experimental results show that the optimized algorithm in this chapter not only requires less computation,but also improves the performance of the algorithm to a certain extent.