Dynamicity-based Crop-Drop: A Context-based 2D Pose Refreshing Algorithm

2D pose estimation plays an important role in many activity recognition applications. However, pose estimation is computationally expensive. In this paper, we propose DCD (Dynamicity based Crop-Drop), a context-based pose refreshing algorithm with controllable information loss and much better processing speed. The main idea is to utilize the correlation between consecutive frames to understand the dynamicity of each individual pose and decide its pose refreshing rate to minimize pose information loss. Full pose estimation is only applied when needed. Our test results show up to twice the speed as compared to that of OpenPose with acceptable information loss when tested on multi-person, multi-dynamicity activities like sitting idle, practicing thai-qi, and dancing.