Research And Implementation Of Multi-object Tracking Algorithm Based On Multi-task Learning

Multi-object tracking,as one of the main components of the machine vision task,has become a current research hotspot in this field.Multi-object tracking is gradually applied to border security,military defense,intelligent transportation and other fields and plays a huge role with the sustainable development of computer vision.Multi-object tracking aims to predict the position of multiple targets in the image sequence of the actual scene and maintain the identity information corresponding to multiple targets.At present,although the target tracking algorithm develops rapidly,it is faced with various difficulties in the actual military battlefield scene,which can be divided into two aspects:(1)Due to the complex and changeable actual battlefield scene,the inaccurate target positioning seriously affects the accuracy of tracking;(2)It’s easy to occur tracking drift phenomenon due to the movement of battlefield target is complex and there are occlusions.Aiming to above problems,in this paper,a multi-object tracking algorithm based on multitask learning is studied.The parameter model is used to share the same semantic expression of different tasks,and a multi-target tracking system suitable for battlefield environment is realized,so as to provide more accurate target positioning and stable ID information.The main research work is summarized as follows:(1)Multi-task collaborative tracking algorithm based on accurate target positioning.In view of the difficulty in extracting complex features in the actual battlefield environment,this algorithm makes use of the spatial pyramid attention and cooperative learning module to enhance the capability of the neural network model,and solve the problem that the target location is inaccurate.In this paper,YOLOv5 will be optimized as the basic algorithm,and the spatial pyramid attention of channel shuffle will be fused to perform multi-scale feature fusion on target features.Furthermore,the channel shuffle module will interact with each other to make the backbone feature extraction network more rich semantic characteristics,so as to obtain more accurate positioning information;secondly,the collaborative learning network learning is used to focus on the relevant semantic information within and between classes,alleviate the contradiction between the detection of multi-task learning network and the imbalance of Re-ID semantic mining,and extract more accurate target detection results to improve the model inference utility.Finally,the results of the experiments indicate that the proposed multi-task cooperative tracking algorithm based on accurate target positioning can effectively alleviate the problem of inaccurate target positioning caused by complex environment.(2)Multi-object tracking algorithm based on anti-drift.In view of the problems of occlusion and complex target movement in the actual battlefield,the multi-dimensional MLP feature enhancement and the motion velocity measurement based on adaptive smoothness coefficient are used to enhance the correlativity and match degree among target and trajectory.In this paper,based on Deep Sort,a multilayered perceptron is proposed to enhance the fusion of multidimension appearance features so as to enhance the appearance feature representation before and after occlusion.When the target correlation similarity is calculated,the motion velocity measurement of the adaptive smoothness coefficient is used to obtain more robust motion information and enhance the correlation degree between the target and the target trajectory.The result that the multi-object tracking algorithm based on anti-drift can track stably in the comp.(3)Test verification and software deployment based on tracking model.To validate the universality of the algorithm and make the tracking system more convenient,different experimental scenarios were designed to verify the proposed tracking algorithm.Meanwhile,a multi-object tracking system based on multi-task learning was designed for practical battlefield applications to test the performance and system functions of the proposed algorithm.