High-speed and low-power VLSI design of global optical flow estimation method for mobile systems

Abstract

Owing to the high complexity of the global optical flow estimation method, it is designed with a VLSI typically for real-time and low-power operation in mobile systems. Most VLSI designs for the global method focus on reducing the excessive external memory access caused by the global and iterative processes. However, to achieve this goal, considerable amounts of resources are used and the throughput is decreased. To address these issues, a multirow-based propagation approach and an efficient VLSI architecture are proposed in this dissertation. To compute the optical flow by using the small internal memory without external memory access, this approach divides an input image horizontally and vertically, and a subimage consisting of multiple rows is processed sequentially. Furthermore, to prevent the accuracy degradation that originates from the boundaries of the divided subimages, boundary conditions are imposed based on the smoothness of the optical flow, which is generally assumed in the global method. In addition, the main equations of the global optical flow method are simplified to boost the processing speed. The experimental results demonstrate that the external memory access is reduced by 96.55% with negligible accuracy degradation. Furthermore, the proposed design uses 80.4%-93.6% less internal memory and achieves 2.2$\times$-29.3$\times$ higher throughput-per-watt compared to state-of-the-art designs.