A systematic Approach to Human Motion Recognition using Deep Learning

Abstract

The application of Human Motion Analysis (HMA) under Computer Vision (CV) is an emerging field which entails various applications such as gait analysis, behavioural cloning and animation of motion, intent detection, etc. For such motion analysis various open source datasets have been created that help analyze motion behaviour. Motion Capture (mocap) files have been used extensively to store motion data and analyze them. Although the weightage of these applications can be huge in modern technology, not much work on human motion recognition has been done using mocap datasets. In this paper, we propose a systematic approach to human motion recognition using software engineering, data analysis and deep learning algorithms. A Deep Learning (DL) model using Gated Recurrent Network (GRU) for the classification of human motion. CMU mocap dataset is used for analyzing motion data and modelling the DL framework. The trained algorithm is tested using accuracy and Mean Absolute Error (MAE) and a user live feed as performance metrics. A 90.1% validation accuracy is obtained on final evaluation.