基于多传感器融合的目标检测方法研究

Research on Object Detection Method Based on Multi-sensor Fusion

为了提高智能汽车对行驶区域障碍物的感知能力,设计了相机图像与激光雷达点云融合的双模中融合模型。选用YOLOv3-tiny作为图像目标检测算法,PointRCNN作为点云目标检测算法,将点云检测获取到的ROI图像与原始图像进行加权融合,在融合后包含障碍物位置信息的图像上进行目标检测。通过与基于图像或基于点云的单模目标检测模型进行比对,在选用合适加权融合系数下双模中融合模型得到了更好的目标检测效果,在KITTI数据集上总类别的mAP@.5:.95上提升了3.3%,在Cyclist障碍物类别的AP@.5:.95上有了5.7%的显著提升。激光雷达点云的引入大大提升了纯视觉目标检测模型在小目标障碍物上的检测能力。

In order to improve the perception ability of intelligent vehicles to obstacles in the driving area,a dual mode fusion model of camera image and laser radar point cloud fusion was designed.YOLOv3-tiny is selected as the image object detection algorithm and PointRCNN is selected as the point cloud object detection algorithm.The ROI image obtained from point cloud detection is weighted fused with the original image and the object detection is performed on the image containing the obstacle position information after fusion.By comparing with the single-mode object detection model based on image or point cloud,it is found that the better object detection effect is obtained when the appropriate weighted fusion coefficient is selected.On the KITTI dataset,the mAP@.5:.95 of the total category has increased by 3.3%and the AP@.5:.95 of the Cyclist obstacle category has increased by 5.7%.The introduction of lidar point cloud greatly improves the detection ability of the pure vision object detection model on small target obstacles.