RGB-Depth Based Gait Data Analysis for Asymmetry Detection

Abstract

In this study, a biometric-based framework is developed to determine human gait abnormality and its severity for application in the healthcare sector for patient status monitoring and pathology diagnosis. The main focus is on detecting the abnormalities, such as asymmetry in gait (antalgic gait). The input data for the proposed framework is provided by non-invasive, no-marker gait data acquisition using the Microsoft Kinect v2 camera. The sample data includes information from a collected database and an externally-sourced database (UPCV). Important gait features are extracted and analyzed through a Dynamic Bayesian Network (DBN) and other traditional classifiers such as Support Vector Machines, K-Nearest-Neighbors and Linear Discriminant Analysis. The DBN-based approach shows better performance and more consistent results than other classifiers. Three separate experiments were conducted. The first experiment examined the feasibility of using lower body joint data (knee angle, ankle angle) to classify gait type. The second experiment examined the use of upper body joint data (shoulder movement) to determine the severity of any abnormalities detected in the first experiment. The third experiment examined how the fusion of both lower body and upper body data impacts the accuracy of decision. Furthermore, a separate experiment was conducted which used the fusion of gait characteristics with facial scores and height measurements to recognize an individual’s identity.