基于YOLOv5神经网络检测模型的动态局部地图构建

Dynamic localized map construction method using the YOLOv5 neural network detection model

在移动机器人自主移动过程中,人的活动会对路径规划结果产生重要的影响。传统SLAM对动态物体的识别和跟踪能力相对有限,动态物体上的特征点容易丢失,且在动态物体占据图像较大范围时,容易造成目标函数的优化方向错误。考虑单个单目相机对行人的位置信息进行检测的场景,通过单目相机获取图像信息,结合YOLOv5目标识别网络来检测图像中人物所占据的像素高度,再通过针孔相机模型还原出图像人物在相机坐标系下的三维坐标。基于YOLOv5网络模型检测出道路区域,运用Canny边缘检测算法计算出道路的边缘轨迹,找到两条轨迹在相机图像上的交汇点。最后根据相机的平面图像结合人物的三维坐标,以及道路轨迹的交汇点和道路宽度信息对图像进行重投影。通过公开数据集测试得到包含行人实时运动情况和道路边界在内的平面地图,可完成150 m深度内的行人位置定位,为快速计算图像的行人的位置深度提供了一种可行方案。

The movement of individuals during the autonomous movement of mobile robots have a significant impact on the effectiveness of path planning.Traditional SLAM methods have limited capabilities in recognizing and tracking dynamic objects.Not only do feature points on dynamic objects tend to be lost easily,but if the dynamic object occupies a large area in the image,it may lead the optimization of the objective function in the wrong direction.This paper addresses the scenario where only a single monocular camera is used to detect the position information of pedestrians.By using image information captured by the monocular camera and the YOLOv5 object detection network,the pixel height occupied by individuals in the image is determined,and then the pinhole camera model is utilized to reconstruct the three-dimensional coordinates of the individuals in the camera coordinate system.The YOLOv5 network model is used to detect road areas,and the Canny edge detection algorithm is applied to calculate the edge trajectory of the road,finding the intersection points of the two trajectories in the camera image.Finally,based on the planar image of the camera,together with the three-dimensional coordinates of the individuals and the intersection points and width information of the road trajectory,the image is reprojected.Experimental testing with publicly available datasets demonstrates the creation of a planar map that includes real-time movement information of pedestrians and road boundaries.The method is capable of accurately estimating the position of pedestrians even at a maximum depth of 150 meters,providing a feasible solution for fast calculation of the depth of pedestrians in images.