Research On Object Tracking And Trajectory Prediction For Autonomous Driving Environment Perception

With the full implementation of the "14th Five-Year Plan",the transformation and upgrading of the automotive industry will usher in new opportunities and new directions for the development of electrification,connectivity and intelligence.Among them,autonomous vehicles occupy an extremely important position.As the specific application of artificial intelligence technology,self-driving cars have achieved impressive results in academia and industry in recent years.Its safe driving experience,intelligent driving mode,and high-efficiency energy-saving benefits are effective in cracking traffic congestion.Many problems such as traffic accidents and air pollution provide high-quality options.At present,the three key technologies of autonomous vehicles mainly include:vehicle environment perception,vehicle route planning behavior intention prediction,and vehicle control.This paper mainly focuses on the visual object tracking and trajectory prediction methods in the key technologies of the autonomous vehicle environment perception stage,focusing on the realization of accurate and robust tracking of vulnerable road users,multiple traffic road participant tracking,and road agent trajectory prediction algorithms.The big breakthrough finally provided key theoretical support for the realization of the construction of an autonomous driving environment perception intelligent system.The main method is to set up visual sensors for self driving vehicles,and object tracking and trajectory prediction are carried out by vehicle autonomously on two elements of road traffic environment and traffic participants,so that they can achieve intelligent information processing ability,efficient decision-making ability and robust and precise control ability in the process of environmental perception.The main work and research results of the paper are as follows(1)Aiming at the problem that Vulnerable Road Users(VRU)are often occluded during the tracking process,a Siamese single object tracking algorithm SiamSC is proposed,which combines spatial features and channel features.The algorithm is divided into two parts:feature extraction and region suggestion classification regression.In the feature extraction stage,in order to reflect the different importance of different positions caused by the occlusion of objects,a spatial attention mechanism is introduced to focus on the important information of the unoccluded parts.In the classification regression stage,the fused features in the classification network are introduced into the channel attention mechanism,and different channels are given different weights,so that the network can better lock the main characteristics of the tracked person,and then perform fast and accurate positioning and tracking.The proposed algorithm has achieved good tracking performance on the single object tracking public dataset and the autonomous driving dataset with Chinese road characteristics.(2)Aiming at the challenges of deformation and scale transformation that often occur in visual single object tracking,a Siamese classification regression single object tracking algorithm SiamMFC based on manifold features is proposed.This algorithm analyzes the previous target tracking algorithm only considers the semantic information of the object and ignores the influence of the rich geometric features of the object on the tracking network.In the network design,the manifold template branch is added to extract the geometric information of the object.At the same time,in the feature extraction stage of the semantic branch,an Anchor point design that requires less calculation and does not need to consider human factors is used to improve the computational efficiency of the network.In the regression stage,the manifold features extracted from the manifold branches are effectively merged,and the object position is accurately regressed.In the experimental stage,in addition to achieving good performance on the single object tracking public dataset,and effectively solving the reduction in tracking accuracy caused by object deformation and scale conversion,the WorldNet data set,GOT10K and UAV123 were also used.Good performance was obtained on the UAV123 dataset taken at high altitude,which met the real-time and robust tracking requirements of autonomous vehicles for vulnerable road users under the challenges of deformation and scale conversion.(3)Aiming at the ID switch problem that often occurs after objects are frequently occluded in multi-object tracking,a resolvable online multi-object tracking algorithm DM-Tracker is proposed.DM-Tracker is essentially a detection-based object tracking algorithm.In the feature extraction stage,the backbone network of multi-feature fusion is used to effectively extract the detection features of the tracked object,and the Anchor-free detection with a small amount of parameters and simple calculation is adopted.The detection accuracy is effectively improved,and high-precision detection results are input for the followup accurate tracking.In the online tracking process,a resolving module is introduced to redistribute features with different weights to the occluded and interactive targets,and focus on whether the unoccluded part of the object and the previous object have the same ID,so as to guide the accurate output of the model.DM-Tracker has achieved good performance on the MOT public benchmark dataset and the KITTI autonomous driving data set.(4)In the process of trajectory prediction,traffic participants,especially pedestrians,will change their trajectories after being affected by external social factors.A state-refined LSTM trajectory prediction algorithm DALATM based on a dual attention model is proposed.DA-LATM considers that the traditional algorithm only uses trajectory information and ignores the influence of scene information in the picture on trajectory prediction,and effectively integrates scene information in the design process to comprehensively predict the trajectory of the road agent.First,the scene image is feature-extracted and sent to the scene attention mechanism module.At the same time,the historical trajectory information is encoded by the state refinement module and then sent to the social attention module to extract important trajectory information.Then the information is effectively fused,and then the fused information is sent to the state refinement module to decode the information and predict the future trajectory.The experimental results on the autonomous driving dataset collected by our country verify the validity of the DA-LATM model.Based on the above research work,this paper starts from different perception tasks,and conducts research based on the visual tracking and trajectory prediction of different research objects.It explores the improvement of the performance of the perception task algorithm under multiple challenges,and then provides the vision for the self-driving vehicle environment perception stage.The task research has laid a certain theoretical and experimental foundation.