Multi-Person 3D Pose Estimation in Mobile Edge Computing Devices for Real-Time Applications

In the last few years, real-time 3D pose estimation from RGB monocular images on Mobile Edge Computing devices has drawn immense attraction due to the ability to estimate, infer and transfer 3D motion and pose in VR, AR, gaming, animation and so on. However, estimating motion under occlusion is very challenging. Though a number of effective and efficient approaches have been proposed to deal with this issue, there is still a demand for robust occlusion-aware multi-person 3D pose and motion estimation under occlusion in real-world scenarios. In this paper, we propose a one-shot occlusion-aware real-time 3D pose estimation and inference approach called RRMP. Our proposed RRMP performs both 2D and 3D pose estimation and is composed of three sequential stages: 1) the residual to render a multi-level perspective for each individual people, 2) the initial stage, and 3) the refinement stages. As our goal is to estimate the pose in real-time for mobile edge computing devices, the RRMP is designed using Depthwise Separable Convolutions (DSCs) that perform with an average of 40 fps in real-time execution. Our extensive results and analysis depict that the proposed RRMP improves the performances of the existing state-of-the-art methods. Our RRMP technique can be deployed into any existing state-of-the-art works for further improving the robustness in terms of occlusion.