Multi-object Tracking Based On Group Proximity Feature Fusion

With the rapid development of deep learning,related applications in the field of computer vision are becoming more and more important.Multi Object Tracking(MOT)is one of the basic applications in the field of computer vision and has great application value in daily life.Such as traffic control,crowd analysis and so on.However,in practices,the task of multi object tracking still faces many challenges: 1)Most existing multi object tracking algorithms mainly extract the appearance features of a single pedestrian,while ignoring the information of other people around the pedestrian;2)Many multi object tracking algorithm can not process the sequence containing time information.Based on the deep learning technology,this paper proposes a method to solve the multi object tracking problem from the above two problems.Aiming at the problem that most existing multi object tracking algorithms ignore the surrounding information of the pedestrian to be tracked,a multi object tracking model based on group proximity feature extraction network is proposed.In daily life scenes and the characteristics of the existing large-scale multi object datasets,in most cases when pedestrians are moving,shopping,running and doing other events,other pedestrians are doing the same or the different events in a certain time and space around them.Aiming at this characteristic,a new feature is proposed,the concepts of group and group proximity feature are defined,the residual network is used as the extraction network,and a variety of loss functions are designed to extract group proximity feature from the input sequence.Experimental results show that the proposed new feature can better describe pedestrian features from three perspectives: global,local and spatial.The model proposed is used to obtain better tracking results in multi object tracking datasets.Although the group proximity feature-based multi object tracking model can extract the group proximity feature of pedestrians well,it cannot process the timing information of the input sequence.To solve this problem,a multi object tracking algorithm model based on multi-feature fusion is proposed.In this model,the Long Short-Term Memory network unit is introduced,the unit is used to store the feature information of the input video sequence,the traditional Long Short-Term Memory network is improved to process the input feature in the way of multiplication,so as to realize the multiplication interaction between the hidden state and the input.Experiments show that the new model that introduces Long Short-Term Memory network units effectively reduces the problem of identity switch caused by factors such as occlusion,and proves that the model can store time information and get better tracking results.To sum up,the multi object tracking based on group proximity feature fusion describes the pedestrian features from three different angles in the spatial dimension.Group proximity feature not only helps to reduce the chaos of global appearance caused by pedestrian joining or leaving the group,but solve the incorrect pedestrian matching problem caused by local feature.At the same time,the Long Short-Term Memory network module is introduced to effectively solve the problem that the proposed feature extraction network can not deal with the time information.The multi object tracking results are more accurate.