Parallel Implementation of Dense Optical Flow Computation on Many-Core Processor

Abstract

Computation of optical flow is a fundamental step in computer vision applications. However, due to its high complexity, it is difficult to compute a high-accuracy optical flow field in real time. This paper proposes a parallel computing approach for fast computation of high-accuracy optical flow field. It is specially designed for Tilera, a typical many-core processor with 36 tiles. By efficiently exploiting the advantages of the mesh architecture of Tilera, and by appropriately handling the parallelism inherent in the optical flow computation, the proposed implemention is able to significantly reduce the computation time while keep a low power consumption. Experiment shows that, for a \(640 {\times } 480\) image, the computation time is only 0.80 seconds per frame. It is 2.56 times faster than on a typical CPU i3-3240 (3.4GHz), and the power consumption as less as 1/6. Experimental results also show that the proposed parallel approach is highly scalable for variable requirements on computation speeds and power consumptions, since it can flexibly selects a proper number of computing cores.