Human Gait Abnormality detection using Low Cost Sensor Technology

Abstract

Detection of gait abnormality is becoming a growing concern in different neurological and musculoskeletal patients group including geriatric population. This paper addresses a method of detecting abnormal gait pattern using deep learning algorithms on depth Images. A low cost Microsoft Kinect v2 sensor is used for capturing the depth images of different subject's gait sequences. A histogram-based technique is applied on depth images to identify the range of depth values for the subject. This method generates segmented depth images and subsequently median filter is used on them to reduce unwanted information. Multiple 2D convolutional neural network (CNN) models are trained on segmented images for pathological gait detection. But these CNN models are only restricted to spatial features. Therefore, we consider 3D-CNN model to include both spatial and temporal features by stacking all the images from a single gait cycle. A statistical technique based on autocorrelation is applied on entire gait sequences for finding the gait period. We achieve a significant detection accuracy of 95% using 3D-CNN model. Performance evaluation of the proposed model is evaluated through standard statistical metrics.