Research On The Localization And Mapping Of Multi-legged Robot Based On Kinect

The multi-legged robot has good flexibility and environmental adaptability,and its structural characteristics of the body redundancy make it able to complete a variety of operations,in the nuclear industry,transportation,fire and other fields.It has good application prospects,with important research value.By installing the sucker on the foot of the multi-footed robot,the multi-foot robot can be moved on the side and the bottom of the bridge,and equipped with the relevant detection equipment to detect the bridge.This is an appropriate choice to realize the bridge detection automation platform.This thesis first designs the parts of the multi-footed robot.Then the kinematics model of the robot is established based on the spin theory,and the inverse kinematics model of the robot is obtained by derivation.Then,the system of self-localization and mapping the environment is designed when the robot moves in an unknown environment.The system includes front and back ends.The front part is image feature extraction,feature matching and pose estimation.The back end includes loopback Detection,pose optimization and mapping.The positioning system uses the Kinect sensor to obtain the environment of the RGB image and depth image.This thesis introduces the feature extraction algorithms of SIFT,SURF and ORB,and uses these feature extraction algorithms to extract the characteristics of the image and compare the appearance.The BruteForce and FLANN feature matching algorithms are introduced,and the nearest neighbor and RANSAC are used to remove mismatching.In using the ICP algorithm to solve the inter-frame motion,we improve the weight of close points to improve the accuracy matching using RANSAC.Considering the effect of Kinect rotation and translational motion on the change of viewing angle scene,we add different weights to the size of inter-frame motion,and select the frame with motion measurement values greater than the preset threshold as the key frame.In order to avoid the vocabulary vector comparison with all frames,this thesis improves the efficiency of loopback detection by constructing vocabulary trees.We then optimize the algorithm by using the Newton Gaussian algorithm by transforming the variables in the local European space in the SLAM problem into manifold and select the octomap library to construct a 3D grid map that can be used directly for navigation algorithms such as A* and D*.At the end of this thesis,we use the fr2 series of open data sets obtained in the industrial hall to validate the SLAM algorithm,and use the average tracking time and absolute trajectory error to evaluate the algorithm.Finally,the multi-legged robot with Kinect sensing is used to verify the algorithm by using the triangular gait in the lab.It is verified that the SLAM algorithm can be applied to the self-localization of multi-foot robots in real time.According to the results of the open set and the real experiment,it can be seen that the SLAM system designed in this thesis can satisfy the needs of multi-foot robot positioning and composition.