| Lower limb rehabilitation exoskeleton as a rehabilitation medical device can help people with lower limb motor dysfunction to achieve walking in different scenarios.Current lower limb exoskeleton robots only focus on certain specific tasks or environments,and cannot meet the needs of patients with lower extremity dysfunction who wear exoskeletons to walk independently in an unknown environment.Therefore,it is necessary to improve the exoskeleton’s understanding of the scene where the environment is located and the identification of the geometric parameters of the environment through the way of environment perception,so as to realize the humanmachine cooperative walking in an uncertain environment.This thesis mainly studies the perception method of exoskeleton in the walking area of daily life environment.The main contents of the thesis are as follows:According to the gait category of the exoskeleton,the walkable area in the daily life environment is divided into four walking scenes of flat ground,slope,stairs and obstacles.For each scene,list the factors and indicators that affect the walking of the exoskeleton,and combine the information of the exoskeleton,barrier-free road design,and the user to obtain the specific index parameters that are feasible in each scenario.Then based on RGB-D camera,a set of walkable area perception system for exoskeleton was designed:(1)For the recognition of exoskeleton walking scenes,this dissertation proposes a scene recognition method based on deep learning semantic segmentation.First,the fuzzy and less semantic information is removed by the Laplacian operator and camera pose;then the pixel-level semantic classification of the image is achieved by the semantic segmentation network;finally,the semantics in the segmentation result are mapped to scene categories,and the walking sense areas are divided by depth data,use the contour detection algorithm to determine the scene type of the sensing area.In the experiment,PSPNet,UPerNet and MobileNetV2 were selected as test networks,and the training set was based on ADE20 K.The results show that the three networks have a correct recognition rate of more than 80% for most scenes.In the small-scale movement,only the stair recognition effect is poor,which will affect the stair segmentation result due to posture and environment.(2)For the recognition of environmental geometric parameters during walking of exoskeleton,this dissertation proposes an environmental parameter recognition method based on camera depth data.In order to reduce the amount of calculation,after converting the camera depth data into point cloud data,a binary map is created from the point cloud depth and height data.The geometric edge information of the environment is obtained by edge detection,and then the upper edge contour with higher information is extracted.All the linear features are extracted by the improved RANSAC algorithm,and the environment parameters are estimated by different strategies for different scenarios.In the experiment,the template with known parameters was used to verify the recognition accuracy of the three scenes of slope,staircase,and cross-obstacle.Except that the width of the side of the obstacle could not be accurately predicted due to the angle of view,the error of the other estimated values was within centimeters.The experiment is also compared with RANSAC and probabilistic hough transform algorithms.The results show that the feature extraction method in this dissertation can guarantee both speed and accuracy of results.The walkable area perception method designed in this dissertation verifies the effectiveness of the overall scheme when walking dynamically outdoors and in stairs in an offline environment.The scene type and the geometric parameters of the environment transmitted by the perception system enrich the environmental information that the exoskeleton lacks,and can effectively improve the decision-making information and safety during walking. |
PDF Full Text Request