Recognition of Abnormal gait patterns using Kinect v2 and Hybrid Deep Learning of Auto-Encoder Feature Extraction Model and Discriminative Model based on RNN Architecture

Abstract

Gait pattern is closely related with health of the human. Gait is an integration of various body functions such as motor function to move and balance the body, sensory function to get information of the surrounding environment and various senses, and cognitive function to make a goal of walking and control it. If one of these functions gets weakened, gait pattern of the man becomes abnormal and it becomes hard to maintain a balance of the body during walking. Therefore, the gait pattern is considered as an important indicator for various diseases. If it is possible to detect some diseases in early stage by analyzing the gait pattern, people can get early treatment before diseases get worse and prevent secondary accidents such as human falls. The purpose of this paper is to recognize the abnormal gait patterns by using Hybrid Deep Learning Algorithm and 3D human skeleton data obtained by using Kinect v2. Various experiments were conducted to verify the effectiveness of the proposed algorithm and to find the best architectural conditions of the model. As a result, the proposed algorithm which uses the features extracted by Deep Layer LSTM Auto-Encoder as the input data of Deep Layer LSTM discriminative model showed the highest performance with 94.69% accuracy in recognition of 9 gait patterns. When the extracted features are used as the input data of the discriminative model, the recognition accuracy is 12.4% higher than the results using the original skeleton data. Therefore, it is verified that the proposed Hybrid Deep Learning algorithm increases the accuracy in the recognition of abnormal gait patterns as well as the robustness to the data and convenience by using the automatic featuring process.