Research On Vehicle Detection,Ranging And Safety Distance Maintenance System Based On Binocular Vision

With the rapid development of driverless technology,vehicle detection,ranging,and safety distance maintenance system based on binocular vision have become a research hotspot nowadays.Compared to monocular ranging,binocular stereo ranging has higher accuracy,which overcomes the disadvantage of low visual ranging accuracy.Compared to other driverless vehicle sensors such as lidar,cameras have a lower cost and the ability to obtain more environmental information for precise control of driverless vehicles.In this paper,the object detection algorithm and binocular stereo ranging algorithm will be deeply studied,and the combination of the two algorithms is applied in the safety distance maintenance system of driverless vehicles,then real vehicle test are conducted to verify the feasibility of the system.The research content of this paper is as follows:The principle of stereo ranging and distortion generation of binocular camera are studied and analyzed.Zhang Zhengyou’s calibration method is adopted to calibrate the camera in combination with the MATLAB camera calibration tool,the internal and external parameters of the camera required for stereo ranging and the distortion coefficient required for distortion correction can be obtained through camera calibration.A proper binocular camera is selected as the sensing equipment,an industrial personal computer as the decision-making unit,and the drive-by-wire vehicle as the executive mechanism to build the hardware architecture of the system.In terms of software architecture,the system is divided into vehicle detection module,parallax ranging module and control execution module based on ROS(Robot Operating System).The vehicle detection model is built based on YOLOv5 s framework.After the vehicle image is input into the model,feature information is initially extracted in the backbone network,then the obtained feature information is sent to the Neck network for multi-scale fusion.Finally,the fused information is sent to the output end for prediction of results.The model is trained with self-collected datasets and network datasets from various road scenarios,a test set is employed to evaluate the performance of the trained model.The results demonstrate that the model exhibits high detection accuracy for vehicle targets and robust performance across diverse scenarios.The parameters obtained from camera calibration are used for image distortion correction and stereo correction to ensure strict alignment of the left and right images,and reduce the workload of stereo matching.The semi-global matching algorithm is adopted to take both accuracy and real-time performance into consideration,the cost is calculated by the original graph and the map graph traversed by sobel operator,the multipath constrained aggregation method is utilized for cost aggregation,the distance information of the target is calculated according to the obtained parallax value.Finally,the target detection algorithm is combined with the stereo matching ranging algorithm to achieve distance calculation of vehicle targets.The safety distance model is designed to assess the current safety status of a vehicle,and the accuracy and feasibility of the overall system are verified by static and dynamic real vehicle experiments.In the static ranging experiment,vehicles at varying distances are measured,and the discrepancies between the actual and detected distances are compared and analyzed,the results show that the system can accurately determine the distance to a vehicle target.In the dynamic real vehicle experiment,the system can realize effective identification and ranging of the vehicle ahead,as well as precise braking when the distance fall below the safe threshold,which verifies the feasibility of the system.