AI-based human whole-body posture-prediction in continuous load reaching/leaving activities

Determining worker's body posture during load handling activities is the first step toward assessing and managing occupational risk of musculoskeletal injuries. Traditional approaches for the measurement of body posture are impractical in real work settings due to the required laboratory setups and occlusion issues. This study aims to develop artificial neural networks (ANNs) to predict full-body 3D continuous posture during load-reaching and load-leaving phases of lifting and lowering activities thus complementing our previous posture prediction ANNs for the load-moving phase (i.e., the lifting activity between load-reaching and load-leaving phases). Using an existing whole-body motion dataset from twenty healthy young novice subjects during 204 load-reaching and load-leaving tasks, four ANNs were developed to estimate body continuous coordinates and segment/joint angles based on task- and subject-specific parameters as inputs. Results indicated that the developed ANNs achieved root-mean-square-errors of <3 cm and <10° for load-reaching and <4 cm and <15° for load-leaving tasks for the whole-body under random hold-out validation. The maximum posture prediction errors were observed at the left side of the body and the prediction errors were larger during the second half of the activities. Compared to prior static posture prediction models, our approach enabled continuous, phase-specific posture prediction thereby improving relevance for ergonomic and biomechanical applications. Although further investigations are required across diverse demographics (e.g., for female, elderly, experienced individuals), the present ANNs represent a step toward more accessible posture prediction tools in occupational settings, potentially reducing data collection demands for ergonomic assessments.