High performance FPGA based optical flow calculation using the census transformation

Abstract

Emerging camera systems in a rising number of cars increase the demand for real-time capable low cost image processing systems. The capability of such a system is demonstrated by an algorithm for the computation of optical flow. In this paper it is shown how the algorithm has to be (re-)designed in order to benefit from platform (FPGA or CPU) specific features. In addition a hardware implementation is presented meeting the real-time requirements and outperforms general purpose CPUs in terms of execution time and power consumption. Timing measurements as well as an analysis on the quality of the resulting optical flow are presented. Although the clock frequency of the FPGA-based solution is nearly 18 times lower, compared to a 1.86 GHz Core2 CPU, results are computed almost twice as fast.