A Lightweight Real-Time 3D Human Pose Estimation Framework for Elderly Monitoring Using Global Optimization and Humanoid Modeling

Abstract

Continuous monitoring of elderly independent living can be important in preventing falls and other accidents, as they are at higher risk. Mobile robots with human pose estimation can make a contribution to the assistance by determining body pose and detect potential hazards in real-time. However, most of the existing pose estimation approaches based on deep learning are unable to correctly estimate uncommon and unexpected human poses. In this work, a lightweight real-time framework for 3D human pose estimation is proposed, which utilizes a detailed humanoid body model, an efficient global optimization algorithm, and an off-the-shelf 2D pose estimation algorithm, and estimates the angle of each body joint with high accuracy. The framework overcomes the depth ambiguity problems that are often encountered during monocular 3D pose estimation by introducing a novel loss function that takes into consideration the constraint of the center of mass with the supporting feet. Furthermore, penalty functions are added to maintain anatomically realistic joint rotations for the estimated pose, leading to more realistic human pose reconstruction.