Research On Object Detection And Tracking Algorithm For Autonomous Driving Scenarios

In the past few years,as China’s modernization process continues to advance,vehicle intelligence,networking and electrification have become a major trend.The research and development of autonomous vehicle driving technology is the focus of attention for major automobile manufacturers.Autonomous driving technology includes environmental perception,autonomous navigation,path planning and decision control technologies.In which,environment perception is the basis of autonomous driving technology.Designing accurate and fast target detection and tracking algorithms can provide assurance for autonomous driving cars to perform subsequent tasks,and can also meet the needs of autonomous driving cars to travel safely on actual roads.This thesis explores the deep learning based autonomous driving environment perception technology,and investigates the lightweight target detection algorithm and target tracking algorithms which are based on convolutional neural networks.In order to overcome the problem of large number of parameters in deep learning objective detection algorithms which are based on convolutional neural networks,this paper proposes the Mobilenetv2＿CA-YOLOv4 lightweight target detection algorithm,which achieves fast and accurate target detection.To solve the problem of partial trajectory loss caused by low detection result utilization during target tracking,based on the lightweight target detection network,the Byte association mechanism is used for target matching and trajectory recovery,thereby achieving accurate target tracking.The algorithm in this paper has achieved excellent results in relevant data sets and real-vehicle tests,fully verifying the effectiveness of the algorithm in this paper,which mainly accomplishes the following work:(1)The Mobilenetv2＿CA-YOLOv4 lightweight object detection algorithm is proposed based on deep learning.The effective lightweight network Mobilenetv2 is used to reconstruct the backbone of the entire target detection network,and a coordinated attention mechanism Coordinate attention(CA)is inserted in it to improve the initial feature extraction ability of the backbone network and solve the problem of insufficient feature extraction;an efficient Mosaic data enhancement method is used to train the network,which solves the problem of insufficient feature extraction for small objects and occluded objects.solves the problem of insufficient feature extraction for small as well as occluded objects;Meanwhile,the algorithm’s loss function is improved based on the CIOU value to address the issue of sensitivity to scale changes during the regression process.Training tests were conducted on the training sets of the classical target detection dataset VOC and the autonomous driving dataset KITTI respectively.,While achieving a lightweight target detection network,the detection accuracy and detection speed are better than those of the comparison algorithm.(2)The object tracking algorithm Byte Track is proposed based on the designed Mobilenetv2＿CA-YOLOv4 object detection algorithm.By dividing the detection boxes obtained from the detection results into highconfidence and low-confidence boxes,and using a secondary correlation matching mechanism to fully utilize the low-confidence detection boxes,the algorithm recovers the target trajectory contained in the low-confidence detection boxes and filters out background detection,solving the problem of lost target trajectories in low-confidence detection boxes.The algorithm is trained on the mixed dataset MOT20 and Crowdhuman,and a comprehensive comparison is made between our tracking algorithm and advanced works on the MOT20 dataset,showing that our tracking algorithm has certain superiority.(3)An experimental platform was built based on the research group’s own experimental vehicle platform,and actual road scene-related images and real-time videos were collected in various different scenarios.For the object detection task,live images from two different scenarios were collected and the proposed detection algorithm was used for detection.For the object tracking task,real-time videos from three different scenarios were collected,and the proposed tracking algorithm was used for tracking.The results of the real-vehicle tests show that the algorithms in this paper all achieve excellent detection and tracking results,and have good generalization performance.