NS-SOF: A non-feature matching approach for sparse optical flow

This paper presents a novel approach to compute sparse optical flow field, which is different from the traditional feature matching methods such as SURF. The approach consists mainly of three novel parts. First, the improved PatchMatch that is tailored to sparse optical flow computation is used to generate NNF quickly. And the NNF is smoothed using a threshold filtering rather than global optimization for low complexity. Next, we use the superpixel method to segment the NNF and choose the representative optical flow for each segment using the statistical filtering. Finally, the outliers that are mistaken for inliers in the previous processing steps are removed using a global statistical filtering in the form of histograms. In the experiment, the approach is evaluated using real datasets provided by the KITTI benchmarks and compared with SURF and ORB. The result shows that our approach can generate more uniform sparse optical flow field with fewer outliers.