Pose estimation and human motion analysis with reinforcement learning

Physiotherapy and health practices at home have entered our lives with development of technology and the spread of digital communication tools. These applications, represent the integration of advanced technologies with the aim of providing more accessible, effective and personalized healthcare services. Nowadays, especially after pandemic conditions, allowing patients to receive treatment comfortably in their own homes and to monitor their health status is an important research topic. Within the scope of thesis study, it is aimed to offer users a low-cost and practical design with a camera-based approach that provides remote physical theraphy by combining health and technology to provide all this need. For this purpose, instructor videos of the movements that users are expected to repeat are shown and they are expected to perform these movements in the most similar way. There are 11 movements selected according to the body function to be exercised in the system, and these movements are presented to the users in grouped formats. The movements performed are analyzed with Mediapipe and the angles extracted specific to the movement are used in the similarity of the instructor/user. These angles are compared with the dynamic time wrapping algorithm, so the similarity of the movements can be evaluated independently of time. The decision of the action to be shown to the user according to the environmental design in the system is made by the policy iteration algorithms, which is one of the reinforcement learning algorithms. A gamified design is presented to the user through reinforcement learning, and the aim is to keep the users's motivation high with the reward/punishment system. The system has been developed to be flexible and suitable for shaping with the suggestion of experts in the field. Although the main application area within the scope of the study is determined as physical therapy, it is possible to apply it in all areas that work with the same logic.