Research On Key Technologies Of Robot Obstacle Avoidance Based On Deep Learning

With the development of industrial automation and artificial intelligence technologies,Robots are widely used in military,geological exploration,home,medical and other fields.The requirements for the use of robots are also also constantly improving in terms of speed,accuracy,and stability.Obstacle avoidance,as a crucial step in robot navigation,has received widespread attention from many researchers in recent years,and a large number of robot obstacle avoidance studies based on machine vision have been carried out.The fusion of super-resolution generation adversarial network and object detection network has emerged in the field of remote sensing object detection,but there almost have no report on robot obstacle avoidance base on the fusion network.Therefore,we initially explore robot obstacle avoidance based on the fusion of super-resolution generation adversarial network and object detection,and carriecd out researches on the main content involved in this thesis:(1)Research on the overall architecture of robot obstacle avoidance and obstacle strategiesFirst,design a deep learning robot obstacle avoidance system based on the robot operating system(ROS);then establish the robot kinematics model and analyze the coordinate transformation;choose the RGB-D-based infrared depth camera Kinect as the robot perception sensor,and analyze the imaging mechanism of the Kinect;finally,make an obstacle avoidance strategy with Dynamic Window Approach(DWA)as a local path planner and object detection network under a global path planner A*glgorithim on a map that has been constructed.(2)Research on object detection based on SRGAN&YOLOv4Building the obstacle detection network by cascading the super resolution generative adversarial network(SRGAN)generator and you only look once version 4(YOLOv4)network:Firstly,analyzing the SRGAN and YOLOv4 network structure and loss functions;we choose a fusion method of cascade to combine SRGAN generator and the YOLOv4 network,training it end-to-end;Then evaluated the fusion network according to the metric of object detection and verified the increasment of accuracy of detection and resolution of images by testing it in real scene images;Finally,the whole netwok was packaged and compiled in the ROS and new cascaded network nodes were established.(3)Robot obstacle avoidance experiment based on deep learningFirstly,we built a platform based on the robot operating system;briefly introducedg ROS development tools and communication mechanisms;Then established the overall navigation obstacle avoidance experiment process and drawed a communication mechanism diagram of robot navigation stack including srgan&yolov4_ros node;And we also analyzed costmap of robot planner and updation of obstacles Layer;finally we verified the feasibility and reliability of the system and analyzed the results of the robot obstacle avoidance experiment by Turtlebot2 in the real environment.In this thesis,Monte Carlo particle filter and odometer message are combined to determine the pose of the robot.On the constructed map,A*algorithm was used to perform global path planning,and a joint obstacle detection system based on SRGAN&YOLOv4 cascaded network and DWA local planner.The cascaded network has a higher value than YOLOv4mAP,it also improves the accuracy of pedestrians detection by 12%and chairs by 1.89%;Since the detection system was added by updating the obstacle information in real time to and local planner and costmap,it helped to reduce the veolicity sampling space of DWA local planner and avoid the behaviour of bypassing the "C" type dynamic obstacle,which was effective to save the total route time.These verified the effectiveness of the joint obstacle avoidance strategy of the cascade network and DWA using deep learning for object detection and super-resolution reconstruction in actual scenes.According to the experiments,the robot obstacle avoidance system designed in this paper has the ability to accurately avoid dynamic obstacles,and also has theoretical significance and practical value.