Fast Human Pose Tracking with a Single Depth Sensor Using Sum of Gaussians Models

What is it about?

We introduce a simple yet effective 3D human pose tracking from a single depth sensor by using the Sum of Gaussians (SoG) models. Both the human body model and the point cloud converted from a depth map are represented by two different SoG models, which allow us to compute and optimize their similarity analytically. We have two main contributions in this work. The first is we extend the SoG-based similarity by integrating two additional terms to enhance the robustness and accuracy of 3D pose tracking. One is a visibility term to handel the incomplete data problem and the other is a continuity term to smooth the motion estimation. Second, we develop a validation and re-initialization strategy to detect and recover tracking failures. Our algorithm is practically promising that neither involves training data nor a detailed mesh or complicated 3D model. The experimental results are impressing and competitive when compared with state-of-the-art algorithms on a benchmark dataset considering the efficiency and simplicity of our method.

Featured Image

Why is it important?

Our algorithm is practically promising that neither involves training data nor a detailed mesh or complicated 3D model.