Evaluation of motion analysis on synthetic and real-world image sequences

Motion analysis is of basic importance for vision-based driver assistance systems. Optical flow is a common image motion representation. This paper provides comparative evaluations of four different optical flow algorithms (called Horn-Schunck, CLG, BBPW, and TV-L₁ in the paper). The algorithms are tested for synthetic and recorded sequences on original image data, on edge maps, and on residual images after applying a smoothing operator. For testing robustness, the algorithms are also analyzed on synthetic sequences with artificial noises (Gaussian blur, Gaussian white noise, or constant brightness changes). A comparative discussion of the algorithms is provided at the end.