A Deep Neural Network-based method for estimation of 3D lifting motions

Abstract

The aim of this study is developing and validating a Deep Neural Network (DNN) based method for 3D pose estimation during lifting. The proposed DNN based method addresses problems associated with marker-based motion capture systems like excessive preparation time, movement obstruction, and controlled environment requirement. Twelve healthy adults participated in a protocol and performed nine lifting tasks with different vertical heights and asymmetry angles. They lifted a crate and placed it on a shelf while being filmed by two camcorders and a synchronized motion capture system, which directly measured their body movement. A DNN with two-stage cascaded structure was designed to estimate subjects’ 3D body pose from images captured by camcorders. Our DNN augmented Hourglass network for monocular 2D pose estimation with a novel 3D pose generator subnetwork, which synthesized information from all available views to predict accurate 3D pose. We validated the results against the marker-based motion capture system as a reference and examined the method performance under different lifting conditions. The average Euclidean distance between the estimated 3D pose and reference (3D pose error) on the whole dataset was 14.72 ± 2.96 mm. Repeated measures ANOVAs showed lifting conditions can affect the method performance e.g. 60° asymmetry angle and shoulder height lifting showed higher 3D pose error compare to other lifting conditions. The results demonstrated the capability of the proposed method for 3D pose estimation with high accuracy and without limitations of marker-based motion capture systems. The proposed method may be utilized as an on-site biomechanical analysis tool.