Environment Perception And Autonomous Motion Realization Of Quadruped Robot Based On Visual Information

The legged mobile robot not only has the motion characteristics of discrete foot support,but also has a variety of gaits such as crawling and diagonal trotting.Compared with other mobile robots,legged robots have stronger terrain adaptability and obstacle crossing performance.Quadruped robot with simple structure and agile movement has always been the focus of research in this field.The ability of sensing environment and autonomous movement is the basis of high dynamic motion of quadruped robot in complex terrain.Pilot following and uneven terrain mapping technology based on visual information are effective methods to improve the maneuverability of quadruped robot.This paper aims at the complex environment perception and autonomous motion needs of quadruped robot,based on the electrically driven SDUQuad-48 quadruped robot,a dual-camera visual perception system was constructed.The pilot stable following method when there are various disturbances in the external environment,2.5 dimensional terrain height map construction and landing adaptation method and non-collision autonomous navigation technology in complex environment are studied,and the researches mentioned above have been validated through simulation experiments and prototype experiments.The following is the specific work content of the paper:(1)A dual-camera visual perception system suitable for autonomous motion of quadruped robot is established.Firstly,the requirements of pilot following,terrain mapping and landing adaptation as well as environmental navigation tasks were analyzed,and the perceptual layout scheme such as angle and field of view for environmental color and depth images collected by robot visual camera was developed.Then,built a visual perception hardware system combining Intel RealSense D435i camera and NVIDIA Jetson Xavier NX development kit.Finally,to achieve the information exchange between multiple machines and control instructions transmission of the system,a system control software architecture based on ROS software development platform and LCM communication mechanism is designed,which establishes the platform foundation for the research of visual assisted autonomous motion tasks.(2)A real-time pilot target detection and following method based on deep learning is studied.Firstly,the quadruped robot detects moving objects in the surrounding environment based on YOLOv3 algorithm.Secondly,a pilot target extraction strategy based on target confidence is developed for illumination variation,personnel crossing,obstacle occlusion and other interference factors.Then,according to the depth distance and angle deviation involving the pilot and the robot,a pilot following controller is designed,and the desired motion instruction is sent to the robot through the LCM communication network.Finally,the experiments are carried out to confirm the efficacy of the pilot detection and following algorithm.(3)A method of terrain perception and foot adaptation assisted by 2.5 dimensional map is studied.Firstly,the terrain point cloud is collected by the tilted depth camera,and a 2.5 dimensional map of environment in the world coordinate system is constructed by transforming point cloud data from the camera coordinate system to the world system.Then,the map cells are discretized,and the expansion and corrosion operations are carried out to reduce the map noise.According to the terrain height in the cells and robot movement information,the terrain passability map based on the potential field algorithm is constructed,which provides reference for the selection of the robot’s mode of passage and landing point.Finally,the effectiveness of the terrain sensing and landing adaptation method is verified by physical prototype experiments.(4)The environmental navigation algorithm based on patrol strategy is studied.Firstly,Intel RealSense T265 camera was used to extract the robot’s plane position coordinates,which was combined with the robot’s own odometer for the space positioning.Secondly,the posture of the robot is updated based on real-time positioning data,and the updated posture information is passed into the elevation mapping algorithm to obtain the dynamic environment map following the robot movement.Then,the position coordinates of obstacles are obtained from the map,and several target track points are given in the obstacle-free area in the map,so as to calculate the distance and angle data between successive waypoints.In addition,the robot motion speed following controller is designed to complete the collision-free autonomous navigation task.Finally,the experimental results confirmed the effectivity of the patrol navigation algorithm.