Research On Single-target Pedestrian Following System For Mobile Robots

Mobile robot technology is a high-tech that integrates a variety of disciplines.It has been developing continuously with the progress of science and technology and widely used in industrial production,logistics transportation and daily life.In response to people’s increasing demand for intelligence,mobile robots need to be able to automatically obtain scene information and plan their own mov ing path,and at the same time,perform dynamic and static obstacle avoidance operations to meet the required functions.In this paper,a four-wheeled mobile robot is used as the platform to study a real-time following system for single-object pedestrians based on visual information and lidar data.Firstly,the real-time object detection algorithm based on visual information is studied,and the detection performance of YOLOv4 algorithm and Tiny YOLOv4 algorithm on the host PC platform and embedded platform on mobile robot is analyzed and compared.According to the detection results and the analysis of the neural network structure of the two algorithms,optimize the network structure by pruning,replacing network layers and adding detection heads respectively,so that the retrained detection network can realize real-time detection of pedestrian objects on embedded devices.The detection network with the best performance results is used as the detector for follow-up studies.Secondly,a real-time object tracking algorithm based on deep and shallow feature fusion is proposed for the condition of single-object following in the system.Calculate the 4-part color histogram of the tracked object and the predicted object to extract shallow features.The detector identifies possible pedestrian object and extracts deep features.Use the shallow features to perform similarity matching of object to judge the reliability of the predicted position.Deep and shallow fusion features are used for more accurate object matching to track the object.The proposed trajectory update strategy can effectively solve the tracking difficulties such as target deformation and short-term occlusion while ensuring the tracking speed.The tracking experiments on La SOT and OTB100 datasets can achieve tracking accuracy performance of 0.581 and 0.453 at real-time tracking speed of 33.64 FPS and35.32 FPS,respectively.Then,aiming at the problems of target position determination and local obstacle avoidance of the mobile robot during the following process,this paper uses the detection method of depth camera and lidar data fusion to obtain the surrounding environment information,so that the system can effectively detect the relative orientation and distance of surrounding obstacles.A moving following method of semi-rigid connection is proposed.The object coordinate position output by the tracking algorithm and the fusion ranging result are used to judge the motion state of the pedestrian object.According to different motion states such as moving away,getting close,and relative static,the underlying control module performs the targeted task to adjust the mobile robot.Furthermore,the problems such as possible target occlusion,target loss and local obstacle avoidance are studied.Finally,the above algorithm is modularized and encapsulated based on the ROS(Robot Operating System),and each functional node and information transmission method are set.The effectiveness of the following system is tested through the Gazebo simulation platform and the real scene in the laboratory.The results show that the mobile robot can effectively follow the pedestrian object according to the motion state.