Research And Implementation Of Obstacle Avoidance Algorithm For Mobile Robot Based On Machine Vision

The key problem to be solved in mobile robot navigation is dynamic obstacle avoidance technology,which will play an indispensable role in the future smart city.In this thesis,the indoor environment is taken as the application scene.Under the premise that the global map is known and the local map is unknown,visual detection,global navigation,and local dynamic obstacle avoidance are studied respectively.To improve the traditional obstacle avoidance method for mobile robots,a local obstacle avoidance algorithm based on the fusion of vision and dynamic window approach(DWA)is proposed.The algorithm research of local dynamic obstacle avoidance based on vision assisted mobile robot is completed,and the design and system construction of the obstacle avoidance algorithm are completed as well.Finally,the related algorithm of this thesis is implemented and verified on the mobile robot platform.The main contents of this thesis are as follows:1.The overall design of obstacle avoidance algorithms for mobile robots is studied,and the modeling,simulation,software design and software platform construction of mobile robots are completed based on ROS(Robot Operating System).Finally,the actual robot is used as a verification platform to transplant and build the obstacle avoidance algorithm based on machine vision in this paper,and finally perform experimental verification.2.Aiming at the disadvantage that the single external sensor lidar cannot fully perceive obstacle information,a vision-based object detection algorithm is improved to improve the robot’s acquisition of mobile obstacle information.In order to meet the real-time requirements on mobile,this thesis improves the object detection algorithm of YOLOv4.Because the original YOLOv4 feature extraction network CSPDark Net53 is more complex and computationally intensive,the network depth is too deep,and the computing power of hardware devices in the real environment is limited to ensure real-time detection,this thesis changes it to a lighter-weight MobileNet V3 network.It ensures the accuracy and improves the detection speed,and is used for real-time detection of indoor moving obstacles in the upper machine of the robot.Although the mean average precision(m AP)of the improved YOLOv4 algorithm is 3% lower than that of the original algorithm,the detection speed is increased by 15 FPS and the detection accuracy basically meets the detection of indoor mobile obstacles.3.Aiming at the problem that the traditional DWA local obstacle avoidance algorithm requires a large amount of sampling and the real-time performance is insufficient,a local obstacle avoidance algorithm based on the fusion of vision and DWA is improved.In the traditional DWA local obstacle avoidance process,the visual detection information of the host computer is integrated,and the simulated motion trajectories obtained after sampling in different speed groups are evaluated.The evaluation function is modified and the visual information is given a certain weight to dynamically select the best obstacle avoidance trajectory.The experimental results prove that the DWA algorithm combined with vision in this thesis is superior to the original algorithm in terms of the average task completion time,the average trajectory length and the average number of rotations,which effectively improves the local dynamic obstacle avoidance ability of the mobile robot.