| Simultaneous Localization and Mapping(SLAM)and path planning are two core technologies for mobile robots.In recent years,the technology of computer vision has developed rapidly.And there are many significant results emerge in many application areas.Visual SLAM is receiving more and more attention from researchers as its application in robotics.Among them,home service robots are the closest to people's lives,but because there are many irregular spaces and potential moving objects in the home environment,visual map building and obstacle avoidance of mobile robots currently have many problems,so the research on visual SLAM and dynamic obstacle avoidance technology of home mobile robots in this thesis has high practical application value.In this thesis,we focus on visual odometry in visual SLAM systems and dynamic obstacle avoidance techniques for mobile robots in the home.The main work in this thesis is as follows:This thesis firstly describes the status of domestic and foreign research on visual SLAM and mobile robot path planning technology,and on this basis leads to the framework of visual SLAM system,and analyzes the camera model,depth camera ranging principle,the principle of feature point method visual odometry and the principle of traditional dynamic obstacle avoidance algorithm.For the problem that the standard feature point method visual odometry has a large error in the estimation of the bit pose in dynamic scenes,this thesis proposes a visual odometry based on dynamic area detection.Firstly,Vibe,a dynamic object detection algorithm with high real-time and small computation,is selected as the basis for detecting dynamic regions,and then the principle of Vibe algorithm for detecting dynamic objects is analyzed,and the problems of its application to mobile robots are pointed out.Vibe is then improved in three areas,namely background model initialization,background model update and "ghost" region processing,to obtain the dynamic region detection algorithm in this thesis.Then,it is fused with the standard feature point method visual odometry framework to obtain the improved visual odometry in this thesis.The experimental results show that the localization accuracy of the SLAM system based on the improved visual odometry in this thesis is improved by up to 68.4% compared to the conventional ORBSLAM2 in an indoor environment where moving objects are present.In order to address the problem that the traditional dynamic obstacle avoidance algorithm does not have a high success rate in the navigation process and is easy to fall into the local optimum,this thesis introduces the idea of reinforcement learning into the problem of dynamic obstacle avoidance.In this thesis,we first analyze the process of reinforcement learning algorithm.Then two typical reinforcement learning algorithms QLearning and Sarsa are compared and analyzed,and Sarsa is chosen as the basis of dynamic obstacle avoidance algorithm in this thesis.Then the environment model for dynamic obstacle avoidance of mobile robot is proposed in this thesis,and the state space,action space of robot and reward function of Sarsa algorithm are redesigned and combined with the prediction of pedestrian trajectory to get the dynamic obstacle avoidance algorithm in this thesis.The experimental results show that the success rate of obstacle avoidance is significantly higher and the overall navigation time is reduced compared to the traditional Dynamic Window Approach algorithm.In order to verify the effectiveness of the improved visual odometry and dynamic obstacle avoidance algorithm of the mobile robot in a real environment,a software and hardware platform for a home mobile robot is designed and built in this thesis.Using this platform to perform navigation experiments in a real environment.The experimental results show that the mobile robot visual SLAM system designed based on the improved algorithm in this thesis is more robust,and the robot navigation success rate is as high as84% in the experimental environment. |
PDF Full Text Request