An acceleration method for correlation-based high-speed object tracking

Abstract We propose an acceleration method for correlation-based object tracking. Correlation-based tracking methods involve an inverse Fourier transform, which is a bottleneck in acceleration. We exploit a trait of high-speed vision to accelerate this computation: namely, we assume that displacements between consecutive frames do not change dramatically within a short frame acquisition interval. By limiting a small region where the displacement is assumed to be, we propose to reduce computational cost in the inverse Fourier transform. We implemented the proposed method in phase-only correlation and conducted experiments using both simulated and real data. We achieved a computation speed around five-times faster than the conventional method without sacrificing accuracy. The proposed method provides as a useful building block in accelerating correlation-based object tracking.