Research On Visual Localization And Motion Planning For Humanoid Robot In Dynamic Environment

The ability of perception and motion is necessary for the humanoid robot to serve humanity.The humanoid robot perceives the environment based on visual information and plans its motion to perform complex tasks,which is an important manifestation of intelligence.Currently,the research on the visual positioning and motion planning for the humanoid robot mainly focused on static environments,and these methods are difficult to be used in the complex dynamic environment of human life.This thesis focused on visual SLAM in dynamic environments and gait planning for humanoid robot climbing stairs.The main parts of this thesis include the improved method of visual odometer in dynamic environments,the method of dynamic object segmentation and extraction,the positioning method based on monocular vision,and the motion planning method of humanoid robot climbing stairs.This thesis firstly introduces the main contents of each part.By summarizing the related works of humanoid robots,visual SLAM,and visual SLAM in dynamic environments,the main researches of this thesis are proposed.Secondly,this thesis presents an improved method for visual odometry based on optical flow,which jointly estimates the camera motion and dynamic segmentation in the direct RGB-D SLAM system.This method focuses on overcoming the interference of dynamic objects by using optical flow residual for improving the ability to perceive dynamic objects in the visual SLAM system.Thirdly,this thesis presents an improved method for visual odometry based on instance segmentation,which segments the object into dynamic and static based on the mobility of the object in dynamic environments using instance segmentation.Fourthly,this thesis presents the gait planning and control method for humanoid robot climbing stairs using improved visual positioning.Finally,a series of experiments were designed for evaluating these proposed methods.The experimental results show that the improved visual odometry increases the accuracy of visual positioning and dynamic segmentation in dynamic environments effectively,and the gait planning and control method makes the humanoid robot successfully climbs multiple steps of a stair.