An adaptable system for RGB-D based human body detection and pose estimation: Incorporating attached props

摘要

One of the biggest challenges of RGB-D posture tracking is separating appendages such as briefcases, trolleys, and backpacks from the human body. Markerless motion tracking relies on segmenting each depth frame to a finite set of body parts. This is achieved via supervised learning by assigning each pixel to a certain body part. The training image set for the supervised learning are usually synthesised using popular motion capture databases and an ensemble of 3D models covering a wide range of anthropometric characteristics. In this paper, we propose a novel method for generating training data of human postures with attached objects. The results have shown a significant increase in body-part classification accuracy for subjects with props from 60% to 94% using the generated image set.