Local Descriptor by Zernike Moments for Real-time Keypoint Matching

摘要

This paper presents a real-time keypoint matchingalgorithm using a local descriptor derived by Zernikemoments. From an input image, we find a set of keypointsby using an existing corner detection algorithm.At each keypoint we extract a fixed size image patchand compute a local descriptor derived by Zernikemoments. The proposed local descriptor is invariant torotation and illumination changes. In order to speedup the computation of Zernike moments, we computethe Zernike basis functions in advance and store themin a set of lookup tables. The matching is performedwith an Approximate Nearest Neighbor (ANN) methodand refined by a RANSAC algorithm. In theexperiments we confirmed that videos of frame size320×240 with the scale, rotation, illumination andeven 3D viewpoint changes are processed at 25~30Hzusing the proposed method. Unlike existing keypointmatching algorithms, our approach also works in realtimefor registering a reference image.