Novel Keypoint Registration for Fast and Robust Pose Detection on Mobile Phones

We present a novel vision-based pose detection method that can be used in mobile AR services. Conventional methods are unable to meet all the requirements such as complexity, robustness and memory consumption for mobile AR services because of their trade-off relationship. In this paper, we propose a novel key point registration approach to solve the problem. Our registration method detects key point candidates and their binary descriptors from a small number of essential training images to improve robustness to changes in viewpoint. The detected features are screened by our two-stage selection method that selects only good features for pose detection. Experimental results demonstrate that our approach both improves the robustness of the conventional method by about 50% and speeds up runtime processing by about 7-10% with small memory consumption.