Research On Object Detection Based On Point Cloud And Image Fusion For Autonomous Driving

Research and development of reliable automatic driving can improve traffic safety and alleviate traffic congestion.For automotive auto drive system,target detection is a very important part of sensing the environment.Therefore,it is of great significance to study the target detection in automatic driving scene.The image obtained by RGB camera has good texture information.The image obtained only by RGB camera lacks spatial depth information,and the detection result is greatly affected by illumination;The point cloud information obtained by lidar is not affected by illumination and has depth information,but the point cloud does not contain texture information.The situation in the automatic driving traffic environment is complex and diverse.When a single sensor is used for target detection,it is easy to miss detection and false detection when the target is far away or partially occluded.In view of the above situation,this paper considers using the characteristics that multiple sensors can complement each other to fuse the image detection results and the point cloud detection results at the decision level.Firstly,considering the high real-time requirements of the application scenario of automatic driving target detection,this paper selects the single-stage detection model YOLOv5 model as the basic network model,analyzes the basic principle of YOLOv5,improves the activation function,introduces the attention mechanism,improves the multi-scale network structure,detects two-dimensional targets for pedestrians,vehicles,etc.,and improves the detection accuracy of small targets to a certain extent.Then,through the analysis and research on the characteristics of the point cloud,the point cloud is projected by using the YOLO3 Dmethod for reference to get an aerial view.The YOLOv2 model in the original YOLO3 D method is replaced by the improved yolov5 network model for detection,which improves the accuracy of target detection.Finally,the detection results of point cloud and image are fused at decision level.The fusion result of the bounding box is determined by comparing the overlap rate and the distance.The fusion of the confidence is based on D-S evidence theory.In this paper,experiments are carried out on KITTI datasets and compared with other literatures.The experimental results show that the improved method proposed in this paper can improve the accuracy and robustness of target detection to a certain extent.