Research On Multi-target Adaptive Motion Tracking Technology Based On Deep Learning

With the development of economy and computer technology,the application field of multi-target tracking technology has become more and more extensive,which is also an important content indispensable for transportation.Installing a driving assistance system in a car can effectively improve vehicle driving safety,and a key technology of the driving assistance system is multi-target tracking.Multi-target tracking technology can obtain external target information through cameras,radars and other equipment,and feed this information back to the driver through the driving assistance system.The application environment of multi-target tracking technology is complex,and it often faces situations such as too similar targets,severe occlusion,camera shake,etc.,which can lead to target tracking failure,loss of target identity,and confusion in recognition.Therefore,it is still a challenging problem to achieve the correct tracking of multiple targets in a more complex environment.For this reason,this paper proposes a multi-target adaptive maneuver tracking algorithm based on deep learning to alleviate the problem of tracking failure under complex environmental conditions.The main research contents of this paper are as follows:(1)Aiming at the problem that the target appearance is too similar,which leads to the exchange of signs,and the wrong identification of the target after occlusion,a re-identification model based on feature aggregation is proposed.First,the channel features in the feature extraction network cannot be effectively used,which leads to the inability of the extracted appearance features to fully express the detection target.An improved feature aggregation network based on the OSA(One-Shot Aggregation)module is proposed,which is spliced by channels.The form of convolution makes full use of channel characteristics.Then use MARS data set to train the proposed improved re-identification network model,and verify it during the training process.The results show that the re-identification model based on feature aggregation is better than the re-identification model without channel feature fusion.(2)Aiming at the problem that the movement of the camera will cause the target trajectory drift,a trajectory correction model based on the ECC(Enhanced Correlation Coefficient)algorithm is proposed on the basis of the deepsort algorithm.When the camera moves,the movement of the target in the image often consists of two parts,one is the movement of the target itself,and the other is the drift of the target caused by the movement of the camera.The motion model in the original deepsort Kalman filter only considers the target’s own motion.To solve this problem,the affine matrix of adjacent images is obtained by introducing the ECC algorithm,and the target trajectory predicted by the Kalman filter is adaptively corrected.(3)The improved re-identification model and the adaptive trajectory correction model based on the ECC algorithm are embedded in the YOLOv3-deepsort algorithm,and the simulation verification is carried out through the MOT16 and KITTI tracking data sets.In the experiment,the algorithm in this paper can alleviate the loss of target identification and the phenomenon of interchange when the target is too similar,occluded,and the camera is moving,and effectively maintains the stability of tracking the target trajectory.