Fast and fluid human pose tracking

Recent advancements in human motion behaviors based on camera images made human motion tracking much more robust than in the past. However, they are still computationally expensive and they do not allow for online reconstruction of the human pose. In this context, this work presents a framework for tracking 2D human pose at high frequency while keeping the robustness of state-of-the-art methods. We achieved this result by combining, by means of Kalman filtering, the recent success of OpenPose, a robust deep learning-based 2D human pose estimation technique, with the fast Lucas-Kanade features matching method. This allows for processing images at different framerates, thus having multirate measurement updates. The frequency of the estimation is kept high by setting the Kalman filter time step. The method was tested on videos of several activities which include both fast and slow motion. The results show an improvement on the reconstruction error, an increased speed of reconstruction, better tracking during fast motion, and the capability to cover loss of tracking from the two measurements. The achieved intraframe (between two available measurements) trajectory estimation frequency was as high as 1 kHz.