Research On Gait Analysis Based On Deep Learning And Plantar Tactile Characteristics

The research of gait analysis is to understand human movement,especially walking.Its main objectives include the measurement of gait function,the detection of irregular behavior and the determination of classification.Gait analysis also helps to develop various tools and devices and provides accurate and effective gait measurement and reliable information for medical experts and researchers.Now,it has become one of the main focuses in the fields of medical diagnosis,security monitoring and sports science.Gait tactile information has attracted more and more attention in the computer vision community due to its rich individual characteristics.Therefore,gait analysis technology based on plantar tactile information has the broad prospect and research value.With the extensive and mature application of deep learning technology in the image field,the research on gait recognition has also made breakthrough progress.With respect to feature matters,the traditional gait analysis methods are all about extracting hand-crafted features,which are inefficient,time-consuming and expensive.However,deep learning models can learn multiple layers of feature hierarchies and automatically build high-level features from low-level ones,which improve the working speed and classification accuracy at the same time.Therefore,the thesis introduces deep learning tools into gait analysis based on plantar tactile characteristics,and focuses on the application of gait analysis in plantar pressure image processing,identity recognition and Parkinson's disease feature quantification.The main work and research results are summarized as follows:(1)In this thesis,a plantar pressure image database of 109 subjects under three different walking speeds of fast,normal and slow speed is established.To a certain extent,the establishment of the database overcomes the problem of small size of the reported plantar pressure image database and has a positive role in promoting the study of gait characteristics.(2)Before classifying the images,the plantar pressure images are firstly pre-processed for registration,and a cascade Convolutional neural network registration method is designed.The model can realize the regression from the difference between the template and the misaligned images to the registration parameters through training.Experimental results show that compared with several traditional methods,the cascade CNN registration method has higher accuracy and shorter computation time.(3)For the classification of plantar pressure images,this thesis designs a gait recognition method based on the DCNN,which takes plantar pressure images as the sole evidence.The DCNN can realize automatic feature extraction and a subsequent built-in classification.Several comparative experiments verify the superior performance of DCNN method.The high recognition accuracy rate(98.47%)on a dataset of 109 subjects proves the feasibility of a trained DCNN for the single-modal gait recognition investigated.(4)In the field of disease diagnosis,taking Parkinson's disease(PD)as an example.A dual-modal attention-enhanced deep learning model is proposed for gait classification and severity level rating of PD patients.The model learns the distinguishing feature from the collected gait Vertical Ground Reaction Force(VGRF)signals,and evaluates the effectiveness of the designed model on four data sets.The experimental results show that the designed model can provide better classification and rating for PD patients,which is very helpful for clinicians to make a better rehabilitation program.