Rehabilitation Training Action Recognition Based On Depth Maps

With the rising of the aging population,the number of people with disabilities and the overall incidence of work-related injuries,the demand for rehabilitation medicine has increased dramatically.The using of rehabilitation robots has replaced some of the physical work.The early emergence of rehabilitation robots can only be used as an implementing agency.With the continuous improvement of human living standards,the demands for mechanical equipment of the patients with limb movement disorders are increasing when rehabilitation training.Based on the serious shortage of the number of rehabilitation therapists and rehabilitation equipment in China at the present stage,the intelligentization of rehabilitation medical equipment can effectively improve this situation and reduce the current work intensity of rehabilitation therapists.This subject focuses on the intelligentization of rehabilitation robots,design and collect the samples of rehabilitation training movement according to the different active joints of patients with physical disabilities,and a variety of algorithms are used to realize the motion recognition.Firstly,the relevant data sets are collected based on specific rehabilitation training actions and the images are preprocess.According to the different stages of rehabilitation,the actions are divided into two parts: single joint motions and multiple joint motions.The collections of rehabilitation training actions were completed by using the Kinect somatosensory equipment.A method is designed that aligning the natural environment depth image with the human area binary image to obtain the depth image of the human body region.In the image preprocessing stage,the morphological processing of the binary image in the human area is performed firstly.Then use the median filter for the depth image of the human body region to perform noise reduction processing to obtain the final depth image of the rehabilitation training action.Secondly,on the basis of depth images,depth motion maps are used as features to classify rehabilitation training actions.The frames of the depth image sequence of the training action are projected on three vertical Cartesian planes,and the frame difference is calculated and the difference image is accumulated to form a depth motion map of theprojection plane.Using bilinear interpolation to unify size of the depth motion map,and using Gaussian normalization to process the data,then PCA is used to reduce the data dimension.Finally,the L2-regularization cooperative expression classifier is used to classify the training action samples.Finally,based on the depth motion map,the local binary pattern is introduced to improve the accuracy of recognition.The depth motion map is divided into blocks,the LBP features of the sub-blocks are extracted,the motion clues represented by the depth motion maps are represented more compactly,and then the statistical histograms of the sub-blocks are performed,and the sub-blocks are Histogram cascading is used to describe the texture features of the entire depth motion map.The PCA is also used to reduce the dimension of the feature vector.In the classification stage,the kernel extreme learning machine(KELM)classifier was used to classify the actions.This dissertation will experiment on the self-made rehabilitation training action data set and MSR Action3 D public data set,and compare with the current mainstream methods to verify the advantages of this algorithm.