Deep neural networks for human motion analysis in biomechanics applications

Human motion analysis is the systematic study of human motion, which is employed for understanding the mechanics of normal and pathological motion, investigating the efficiency of treatments, and proposing effective rehabilitation exercises. To analyze human motion, accurate kinematics data should be extracted using motion capture systems. The established state-of-the-art method for human motion capture in biomechanics applications is using marker-based systems, which are expensive to setup, time-consuming in process, and require controlled environment. As a result, during the past decades, researches on marker-less human motion capture have gained increasing interest. In this thesis, by utilizing advances in computer vision and machine learning techniques, in particular, Deep Neural Networks (DNNs), we propose novel marker-less human motion capture methods and explore their applicability for two biomechanics applications. In the first study, we design and implement a marker-less system for detecting non-ergonomic movements in the workplaces with the aim of preventing injury risks and training workers on proper techniques. Our proposed system takes the workers’ videos as the input and estimates their 3D body pose using a DNN. Then, critical joint loads are calculated from resulting 3D body pose using inverse dynamics technique and are compared with human body capacity to predict potential injury risks. Results demonstrate high accuracy, which is comparable with marker-based motion capture systems. Moreover, it addresses marker-based motion capture system limitations by eliminating the need for controlled environment and attaching markers onto the subject body. In the second study, we design and implement another marker-less system for detecting gait abnormalities of patients and elderly people with the aim of early disease diagnosis and proposing suitable treatments in a timely manner. We propose a computationally efficient DNN to estimate 3D body pose from input videos and then classify the results into predefined pathology groups. Results demonstrate high classification accuracy and rare false positive and false negative rates. Since the system uses digital cameras as the only required equipment, it can be employed in patients and elderly people domestic environments for consistent health monitoring and early detection of gait alterations or assessing treatment outcomes progress. The ultimate goal of this study is providing a tool for Ambient Assisted Living. Ambient Assisted Living is the use of technology, in particular Artificial Intelligence, in people’s daily life with the goal of recognizing actions and detecting events within an environment. It enables a remote health monitoring of patients with chronic conditions and senior adults and helps them live independently for as long as possible.Ph.D.Includes bibliographical reference