Research On Spatial Temporal Attention Based Graph Convolutional Networks For Pathological Gait Auxiliary Diagnosis Algorithm

Robust walking is a task that requires coordination among multiple organ systems in the human body.If any of these organs undergo pathological changes,corresponding changes will occur in the human gait posture.Traditional Chinese medicine and modern Western medicine have both found that early identification of pathological gait is of great significance for disease prevention and treatment.However,accurately classifying pathological gait is a difficult task in medical diagnosis because it requires professional medical knowledge and its conclusions typically depend on the doctor’s personal experience,which can lead to subjective results.This thesis proposes a spatiotemporal attention graph convolutional pathological gait diagnosis algorithm based on 2D human skeletal point sequences to achieve objective,accurate,and fast pathological gait recognition.Firstly,in response to the problem of easily disturbed inputs in current algorithms,this paper designs and implements an automatic gait cycle detection algorithm,which can extract motion information from various joints of the human body and perform denoising.At the same time,the discrete Fourier transform is used to calculate the motion cycle of different types of pathological gait,and the model input is optimized to reduce calculation and improve accuracy.To improve the effectiveness of pathological gait recognition,this paper optimizes the traditional spatiotemporal graph convolutional network using attention mechanisms.Specifically,this study uses a global attention mechanism to optimize the traditional spatiotemporal graph convolutional network,weighting it in the time dimension to help the algorithm extract more effective gait recognition features.Inspired by self-attention mechanisms,this study improves the human skeletal point connection structure,adopts a global connection idea,and uses attention mechanisms to weight different body parts(such as arm and leg movements,head movements,and shoulder movements)on spatial dependency relationships to optimize feature extraction of the model.Through ablation and comparative experiments on public and our datasets,the proposed algorithm achieved an accuracy of 96.54% on our dataset,and 98.78% on the GIST pathological gait public dataset,which is 4.64% higher than ST-GCN.The above experiments demonstrate the accuracy and effectiveness of the Spatial Temporal Attention Based Graph Convolutional Networks for Pathological Gait Auxiliary Diagnosis Algorithm.