Insole-based Real-time Gait Analysis: Feature Extraction and Classification

Abstract

Gait assessment relies on clinical tools based on observation by trained staffs who give a subjective opinion. Objective gait analysis via motion capture systems (e.g. Qualisys) have limited availability as they are laboratory based and require complex equipment. A low-cost user-friendly Inertial Measurement Units (IMUs) embedded insole and an Android App based personalized gait analysis system is developed for uses in home or clinics. Accelerometer and gyroscope synchronous data are collected from both right and left legs for 10 young and 10 older adults a period of 100 consecutive days. We propose an automatic gait features extraction method, real-time visualization and age-groups classification. Accuracy of stride detection method is 100% for young. Accuracy for older adults is 91% for right and 88% for left leg. Convolutional neural networks (CNNs) are used to extract features from gait data and are combined with long short-term memory (LSTM) to exploit the time information between features. This is evaluated empirically using traditional classification and deep learning techniques (CNN+LSTM RNN) regardless of feature engineering. Accuracy to classify young and older adults with CNN-LSTM, NB, SVM and J48 is 100%. Our insole-based gait analysis automatically interprets the gait features and users can monitor their gait at home using our simple visualization tool that allows widespread home-based diagnosis and management of gait abnormalities and rehabilitation.