Research On Multi-target Tracking Method Based On Deep Learning

Video multi-target tracking is one of the important technologies in the field of computer vision.The key point of this technology is to continuously track multiple objects of interest in the video.It is often used in the fields of autonomous driving,video surveillance and intelligent transportation.How to use deep learning to improve the performance of multi-object tracking has become one of the research hotspots in the field of computer vision.At present,there are three difficulties in multi-target tracking technology:firstly,target location,that is,how to detect the target of interest to be tracked;Secondly,how to predict the motion of the target of interest,that is,the identity information of the target is transmitted frame by frame without loss;The third is how to track the target when it is blocked by obstacles.This paper conducts targeted studies on the above three problems,mainly including deep learning-based target detection technology,target motion state estimation,feature selection and feature fusion method.The specific research contents are as follows:(1)Research on target detection technology based on deep learning.This paper adopts the first stage of strategy design without anchor point box.The target detection model is used to predict the central point thermal map instead of the target regression mechanism from the preset anchor point frame.Backbone Network using the improved Hourglass Network depth feature extracting,conducted on a large scale maps the target classification of regression prediction,to prevent a single figure scale feature information is missing,Network architecture with the method of jump layer dense connectivity information aggregation space and scale,in order to ensure that the last layer characteristics can be predicted pixel level.At the same time,Re-ID depth apparent feature extraction branch is added to the model output to extract the re-recognition features of corresponding targets,so as to realize the sharing of convolutional features.(2)Motion state estimation.After detecting the target to be tracked,it is necessary to carry out motion prediction for the target.In this paper,the extended Kalman filter(EKF)algorithm is used to establish the quadratic motion model with constant rotation rate and speed,which is used to predict and update the motion state of the target on the twodimensional image.The effectiveness of the modeling of EKF is verified by simulation experiment and actual tracking results.(3)Research on multi-feature fusion method.After completing the detection and motion state estimation of the target to be tracked,it is necessary to update the tracking state with the detection information of the current frame.To associate the two of the same target,similarity should be measured by using features.In order to describe the target comprehensively,HSV color histogram feature,Re-ID depth apparent feature and motion feature are selected in this paper,and chi-square distance,cosine distance and Mahalanobis distance are respectively used to comprehensively measure the similarity of tracking trajectory.At the same time,in order to make the multi-target tracking algorithm in this paper adapt to multiple tracking scenarios,the three features are weighted and fused,and the weight configuration with the best performance is designed for common tracking scenarios as a reference.According to the results of similarity measurement,the KM binary graph matching algorithm is used to detect and track the track.In order to deal with the occlusion problem,a state update algorithm based on the survival is designed to save the track of the occluding target within a certain period of time without losing,and ensure the effective retrieval of the target when the target reappears.