Camera-Based Multiple Object Tracking For Autonomous Vehicles

With the rapid development of computing power and storage capacity,the artificial intelligence technology has ushered in another peak.As an important foothold of artificial intelligence,autonomous vehicles have also received more and more attention,and they represent the future direction of the automotive industry.Environment perception technology plays the role of eyes of autonomous vehicles.The safety of unmanned vehicles can be guaranteed only if they accurately obtain information about dynamic obstacles of the surrounding environment.The complexity and dynamics of the actual traffic environment put forward higher requirements for the dynamic obstacle perception ability of autonomous vehicles.This thesis has carried out related research on the autonomous vehicles’ multi-target tracking technology.First of all,to solve the problem of the crowded pedestrians that often cross and occlude each other in a traffic environment,a post-processing technology for target detection based on a prior knowledge of location is proposed to solve the positioning issue in the crowded scenes.The head-body detection result data is post-processed by using the clustering prior knowledge of the two associated target anchor frames of the head-body to enhance the pedestrian detection ability under occlusion.The results on the Crowd Human data set show that this method can achieve better pedestrian detection effect.Secondly,the problem of pedestrian re-identification in the multiple object tracking scene is explored.For the multi-object tracking scenes,it is necessary to distinguish the characteristics of multiple pedestrian instances in the current video stream,and a reidentification neural network module that can achieve cross-instance regression attention is designed.By associating image features in different instances during the feature extraction process,the network can dynamically adjust its feature weights and achieve better pedestrian measurement feature extraction.Experiments on various pedestrian re-identification data sets further illustrate the effectiveness of the attention extraction method.Finally,the issue of multi-target tracking under occlusion conditions is studied.Aiming at the problem for strong assumptions about the spatial consistency of the adjacent frames in the video stream of Tracktor framework,the Kalman filter is introduced for better object motion prediction.And an occlusion adaptive target matching metric calculation method is proposed,which solves the problem of poor performance of Tracktor under crowded occlusion conditions and improves the accuracy of multi-object tracking.Through experiments on the MOT17 data set,the effectiveness of the proposed method is verified.