DeepPose: An Integrated Deep Learning Model for Posture Detection Using Image and Skeletal Data

Abstract

Identifying human actions and postures presents significant challenges for computerized systems. The categorization of these tasks holds particular relevance in the fields of health and robotics. Leveraging artificial intelligence technologies, it becomes feasible to define and classify recurring physical movements and postures accurately. Proper posture is further essential for the rehabilitation of patients because it affects the effectiveness of exercise training. Unfortunately, patients fail to follow the correct sequence when performing the exercises. To pursue the problem, a new method is proposed for posture recognition and pose estimation that does not require wearable devices. The proposed model utilizes 2D coordinates derived from the 2D poses as inputs with 18 joints of the human body as key points, along with an image dataset, to accurately classify various postures. This study involved training a custom CNN named DeepPose using both image and keypoint datasets and conducting a comparative analysis with the performance of two other pre-trained models. The result shows that the proposed model with the keypoint dataset outperforms over image dataset.