Efficient keypoint matching for robot vision using GPUs

Finding correspondences between two images is a fundamental task in computer vision. Focusing on robot vision this task has to be computed as fast as possible allowing further computations to finish within frametime. Given this viewpoint the performance of the correspondence task determines the robot's visual abilities. In this paper we present a framework for robot vision allowing this task to be solved at rates of up to 200 frames per second. These high framerates are achievable by consistently optimizing a SIMD implementation for robot vision use cases. The three steps feature extraction, feature description and feature matching are explained in detail. Using a dense calculation of haar wavelet responses based on integral images, the rank of the structure tensor is approximated to determine keypoints with subpixel accuracy. Due to scaling the wavelets themselves, no prior image smoothing or scaling is required. For keypoint matching these dense haar wavelet responses are reused to build image descriptors. To avoid heavyweight image descriptor comparisons, an efficient prematching step based on simple bitmask operations is introduced. A CUDA-implementation is available for download.