Research And Implementation Of Object Detection Algorithm Based On Fusion Of Vehicle-mounted LiDAR And Vision

In the domains of self-driving and intelligent transportation,multi-sensor fusion perception is a necessary means to achieve high-precision,robust autonomous navigation and environment perception Because of the distinct perceptual characteristics and limitations of various types of sensors,it is not feasible for a single sensor to fully capture complex and dynamic scenes.Therefore,multi-sensor fusion is required to acquire more precise and allinclusive environmental information.Sensor fusion technology can enhance the accuracy of tasks such as target detection,as well as improve the robustness and reliability of the entire system,tracking,and localization,but also enhance the system’s robustness and adaptability,and strengthen the capabilities of autonomous navigation and environment perception.The object detection method based on image vision,point cloud and multi-sensor fusion is studied in this paper.Based on a comprehensive review of related research both domestically and abroad,this article delves into two widely used network structures for object detection,YOLOv3 and SECOND,proposes improvements to these two detection methods,and introduces a multi-sensor fusion-based object detection method.The main contents of this study are as follows:1.A refined YOLOv3 algorithm is suggested,which is built upon the YOLOv3 object detection algorithm and incorporates several key technologies to enhance its performance.These technologies include K-means clustering algorithm for re-calculating anchor boxes based on the KITTI dataset,SENET attention mechanism,SPP,GIo U,and Mish activation function.The application of these optimization methods has significantly improved the performance of the algorithm in object detection.Finally,a modified YOLOv3 algorithm is proposed,which is developed based on the YOLOv3 object detection algorithm.2.This paper introduces an enhanced 3D object detection approach based on the SECOND algorithm.The paper first introduces the original SECOND algorithm and CBAM model,and then analyzes the application of CBAM model in SECOND algorithm and proposes further improvements.Finally,the improved CBAM model is inserted between the sparse convolution and RPN network in the SECOND network structure to enhance the 3D object detection algorithm’s performance without compromising its real-time capability.The effectiveness of the proposed algorithm is evaluated on the KITTI dataset by submitting the results to the KITTI server.3.This paper proposes a decision-level information fusion-based object detection method for target detection in car environments,which integrates an improved 2D and 3D object detection method based on CLOCS fusion network,combining the strengths of SECOND and YOLOv3 algorithms.The proposed method achieves accurate detection of objects in the surrounding environment of vehicles through the decision-level information fusion framework of the CLOCS network.Finally,the proposed decision-level object detection method is tested on the KITTI dataset.4.A platform for autonomous driving of BYD Qin EV was built,in which the hardware part consisted of the BYD Qin EV new energy vehicle and various sensors including onboard cameras,Li DAR,millimeter-wave radar,GPS,and inertial measurement unit.The software algorithm framework for the BYD Qin EV autonomous driving perception platform was built on the Robot Operating System(ROS),and data collection,storage,and processing were performed through the perception platform for autonomous driving of BYD Qin EV.In this article,target detection techniques for different sensors were analyzed,including 2D and 3D object detection as well as fusion-based target detection techniques using monocular camera and Li DAR.Actual vehicle experiments were carried out to assess the precision and practicality of these techniques in recognizing vehicles,pedestrians,and cyclists,and the detection outcomes were analyzed statistically.