Research On Multi-target Tracking Algorithm Based On Multi-scale High-resolution Network

As an important branch of computer vision research,multi-target detection and tracking has great application prospects in many fields such as large-volume personnel monitoring,unmanned driving technology,and modern weaponry.However,in actual scenes,it still faces many problems such as occlusion,similar target interference and re-identification.The detection-based multi-target tracking algorithm divides the detection and tracking into two independent subtasks,uses an excellent detector to balance the quality of different observation inputs,and can also adjust the target feature block according to the detected bounding box to deal with the scale change of the target.However,both the detection network and the data association network require a lot of inference time,the model parameters cannot be shared,and the tracking speed is limited.This paper studies an anchorless detection method based on Multi-scale Fusion High-resolution Network(MFHRNet),and implements an end-to-end multi-target tracking algorithm by adding Re ID feature extraction branch.The work is as follows:First,an anchorless detection algorithm based on MF-HRNet network is proposed.Through the interpolation and fusion of different branches of the traditional high-resolution network,a multi-scale high-resolution feature extraction network is realized to enhance the expressive ability of the model.Use the Center Net method to obtain the heat map representation of the input image,and convert the detection problem into a target center point estimation problem.The peak point of the heat map is the center point of the target.The width and height information and offset are used to calculate the frame position to improve the detection of the algorithm.Precision and recall rate.Second,in order to solve the problem of target loss caused by occlusion in crowded scenes,a dual matching attention network is proposed in the data association stage,which combines the characteristic information of the spatial attention network and the timing information of the time attention network to improve the accuracy of the association.The feature pool is added to retain the target features whose trajectory disappears in the fixed frame,and the trajectory matching of the newly emerging target is performed to improve the model drift problem caused by missed detection.Then,the Re-ID feature extraction branch is added to the MF-HRNet network to realize an end-to-end multi-target tracking algorithm that shares detection and tracking parameters,reducing the network's reasoning time.In terms of enhancing the ability of the network to re-identify,adding a hole convolution to improve the receptive field of the model and reduce the loss of information in the process of network transmission.In the output stage,deconvolution is used to increase the resolution of features and improve the small target tracking ability of the model.Finally,the algorithms involved in the article are tested on the public data set MOT17 and compared with the current mainstream tracking algorithms.The experimental results show that the proposed algorithm has better tracking performance in crowded scenes,can improve the problems of occlusion,re-recognition and small target tracking,and has certain advantages in real-time.