Research On Vision Assisted Technology For Exoskeleton Robot

Lower exoskeleton robots can significantly improve the motor function of the lower limb disability and semi-disabled people,allowing them to stand up again.As a wearable body function amplifying device,it pursues the intelligence of robots and the comfort of human-machine coordination.With the gradual deepening of research,the research of exoskeleton will focus on intelligence,so that it can get out of laboratory and be helpful in daily life.The main work of this project is to study two key assistive technologies based on machine vision to realize the autonomous understanding of the physical support re-lationship of the exoskeleton robot in the environment,and to plan the footsteps ac-cording to the current terrain from the starting state to the target state.At present,scene understanding technologies is relatively mature,including scene classification,semantic segmentation and object recognition.But there are few studies on contextual physical relationship reasoning.The common scene understanding technologies focus on the semantic understanding,and it is difficult to explain the relationship between objects.However,the understanding of physical relationships are important for the robot to interact with the scene.In addition,current exoskeleton robots usually use pre-programmed control methods and human-computer interaction control methods.These methods are not intelligent enough,the operation is complicated,or the walking mode is fixed.Most of them stay in the laboratory and can only be used for fixing.occasion.At present,few researchers studied the footprint planning of exoskeleton robots in un-known environment with obstacles through visual information,which is important for achieving autonomous walking of exoskeleton robots.In order to promote the interaction between exoskeleton robots and the environ-ment and improve the intelligence of exoskeleton robots,this paper focuses on visual aid technology for exoskeleton robots,including the following two parts:1.Inferring the support relationship between objects in the scene understanding,re-annotating the pixel-by-pixel support relationship of the NYU V2 data set,and cit-ing a CRF-based probability and statistical model to infer the support structure between objects.Semantic segmentation network is innovatively defined as multi-class support relationship,and then combined with probability map model CRF-RNN to improve sup-port power inference,realizing the automatic learning of a priori structural information through end-to-end deep learning network,and evaluates the physical relationship in the scene.2.Based on the navigation of the lower extremity exoskeleton robot in the ob-stacle environment,the search-based footprint planning algorithm is studied,and the algorithm is improved for real-time performance required by the exoskeleton robot.The innovative footstep planning algorithm based on Bessel channel is proposed.A search space that satisfies kinematic constraints,start state constraints,and target state constraints is constructed,and the number of states of the heuristic function expansion is reduced without increasing the path loss.The experimental results show that the first visual aid technology(scenario under-standing based on support force analysis)can effectively infer the support relations be-tween objects in the scene.The second visual aid technology(footprint planning based on Bessel channel)can realize the footprint planning of the exoskeleton robot in the obstacle environment,and guarantees the real-time acquirement in certain enviroment.