Describing Motion in Computer Images’ Stream: Optical Flow

This chapter gives a review of optical flow and evaluates the performance of the benchmark optical flow techniques on both synthetic and real image sequences. In the comparison, the results show that Horn-Schunck is not good at handling large motion. For block matching, the block size is critical to the performance: if the block size is too small it might not able to integrate all the area, and if the block size is too large it will lead the window to include the motion belonging to other group (Sun, 2013). Farneback (2013) is not good at preserving the motion boundaries. DeepFlow appears to outperform the other techniques in most scenarios (Weinzaepfel et al., 2013).