Research Of People Detection And Tracking Based On 2D Laser Scanner

2D laser scanner is a widely installed distance sensor,which is often used for Simultaneous Localization And Mapping,navigation and obstacle avoidance.Due to its wide perception angle and accurate distance awareness,2D laser scanner based people detection and tracking has become an import research direction in robot perception recently.Leg detection is widely considered as a first-step of people detection.However,the existing leg detectors ignore the inevitable noise and the multi-scale characteristics of the laser scan.It leads to many unreliable features of point cloud and further degrading the performance of leg detection.In the people tracking task,motion model,e.g.,Kalman filter,is usually used to generate the prediction position of the next time.However,these existing methods ignore the influence of people interaction,which may predict inaccurate position and lead to wrong people association.Furthermore,it can degrade the performance of people tracking.To solve those two problems,this paper researches on 2D laser-based people detection and tracking method.First of all,in the people detection subtask,for the outliers in point cloud,this paper designs an effective and simple point cloud filtering method,and optimizes the breakpoint-based segmentation strategy,so as to make the segmentation more complete.In order to solve the noise of point cloud and multi-scale characteristics,a multi-scale adaptive-switch random forest is proposed in this paper.Among them,the adaptive decision tree is proposed to solve the noise of the point cloud.Multi-scale random forest structure is designed to learn the multi-scale characteristics of features.Finally,the high-confidence legs are associated with each other with constraints,and the minimum cost match is solved to detect people.Then,in the people tracking subtask,due to the inaccurate position predicted by the motion model,this paper introduces a social-based people trajectory prediction model.Then a motion-social fused position prediction model is introduced,which can fuse the predicted position into a Position Probability Graph.Finally,a Position Probability Graph and Local Occupied Map based people association strategy is proposed to track people.Finally,several quantitative and qualitative experiments are performed on Moving Legs and People Tracking datasets,and the experimental results are well analyzed.The experimental results illustrate that this method has achieved considerable gain in all evaluation metrics.