FPGA Implementation of Optical Flow Algorithm Based on Cost Aggregation

The computational complexity of the optical flow estimation is very high, and many hardware systems have been proposed. In these systems, Lucas-Kanade, tensor-based, and phase-based method have been widely used. Census-transform, which is widely used in the stereo vision systems, was also implemented in several FPGA systems. In these systems, only one clock cycle is required for calculating one flow as their throughput, and their processing speed is fast enough for real-time processing of high resolution images. GPUs have also been used, and it was reported that the acceleration by FPGAs and GPUs is comparable[1][2]. The main problem in these systems is their low accuracy. The methods described above show high accuracy for the regions with high changes of brightness, but show poor results for uniform regions. This is the common problem with the stereo vision, and the approaches used in the stereo vision can be applied to the optical flow estimation. In this paper, we extend a cost aggregation algorithm[3] for the optical flow estimation, and implement it on FPGA.