Path Planning for Mobile Robot Based on RGB-D SLAM and Pedestrian Trajectory Prediction

Abstract

This paper implements the construction of three-dimensional point cloud map based on RGB-D SLAM, and takes the three-dimensional point cloud information output by SLAM system as the input of Octomap, generates octree map and performs three-dimensional projection transformation, which converts point cloud map into two-dimensional raster map for path planning research. In order to solve the problem of secondary obstacle avoidance and the problem of emergency stop caused by no local optimal solution in local path planning, a new path planning algorithm for mobile robot based on dynamic object trajectory prediction is presented, and the best path is selected by combining the reinforcement learning algorithm, Sarsa, to avoid dynamic obstacles effectively. On the basis of using RGB-D camera to locate pedestrians in real time, Kalman filter algorithm is used to predict the global coordinates of pedestrians in the next moment. Then a new reward and punishment mechanism is designed to realize the dynamic obstacle avoidance based on the improved Sarsa algorithm, so that the mobile robot can leave the radiation circle of pedestrian prediction coordinates as soon as possible and avoid the pedestrian walking path.