Pedestrian Detection System For Near-ground Mobile Platform Based On Human Leg Detection

With the gradual improvement of living standards and technological levels,the use of mobile robots penetrated into grassroots life.Mobile robots are capable of perception,planning and decision-making.The pedestrian detection is a basic and important demand for other tasks.Because pedestrians are non-rigid,indoor environment is not the same and height of pedestrians are very uncertain,the detection algorithm needs to be more flexible.Mobile robots can rely on the equipped depth camera to acquire RGBD data at a higher frequency for environmental awareness.For near-ground mobile platforms such as Kobuki,the pedestrian detection algorithm cannot be used on such platforms due to the sensor field of view.This paper proposes a detection algorithm based on human leg detection applied to near-ground platforms.Compared with the upper body information,the mobile platform pays more attention to the position and information of the pedestrian's lower body.In consideration of the data structure characteristics of 2D images and 3D point clouds,this paper uses different algorithms to process them.The two-dimensional image information structure is more comprehensive and complete,so this paper first uses the improved version of YOLOv3-tiny to detect the human legs in the RGB image.The result of the detection of the bounding box and the D channel data can be combined to obtain a cone angle of view.The number of point clouds in the shape perspective is small but most of them are related to the legs of the pedestrian.This not only reduces the computing pressure of the point cloud segmentation network,but also increases accuracy of the segmentation network.Then Point Net segmentation network is used to the point cloud to obtain pedestrian data.The pedestrian detection algorithm based on human leg detection proposed in this paper has great performance in both accuracy and real-time performance.In the data set produced in this paper,the real-time performance of the algorithm can reach 47.89 fps,and the accuracy of point cloud segmentation can reach 70.27%.