GPU Acceleration in a Visual Servo System

In this paper we present our novel work of using the Graphic Processing Unit (GPU) to improve the performance of a homography-based visual servo system. We propose a GPU accelerated Efficient Second-order Minimization (GPU-ESM) algorithm to ensure a fast and stable homography solution, approximately 20 times faster than its CPU implementation. To enhance the system stability, we adopt a GPU accelerated Scale Invariant Feature Transform (SIFT) algorithm to deal with those cases where GPU-ESM algorithm performs poor, such as large image differences, occlusion and so on. The combination of both GPU accelerated algorithms is described in detail. The effectiveness of our GPU accelerated system is evaluated with experimental data. The key optimization techniques in our GPU applications are presented as a reference for other researchers.