RGB-D human posture analysis for ergonomie studies using deep convolutional neural network

Human posture analysis is a task of utmost importance for several disciplines. For ergonomists, extracting postural information such as joint angles is necessary to evaluating ergonomie assessment metrics. This allows the early identification of potential work-related musculoskeletal disorders in manufacturing industries, and thus providing adequate interventions. In this paper, we present a holistic posture analysis system that estimates body joint angles from an input depth image. The proposed method utilizes the low cost Kinect sensor for data acquisition and a deep convolutional neural network model for joint angles regression. Further, we rely on learning from synthetic training images to allow simulating several physical tasks by different workers and obtain a highly generalizable learning model. The corresponding ground truth joint angles have been generated using a novel inverse kinematic stage. The proposed method achieves high joint angels prediction rate by recording an average MAE of 4.67 deg and RMSE of 6.64 deg.