Monocular optical flow navigation using sparse SURF flow with multi-layer bucketing screener

This paper presents an approach capable of recovering the trajectory from the image sequence of a single camera based on sparse SURF flow with multi-layer bucketing screener. The main novelty of the proposed approach is that we designed a multi-layer screener to screen good features detected by SURF algorithm. The bucketing scheme firstly divides the image into several buckets. For each bucket, both mismatched and redundant feature pairs will be rejected so that the accuracy of SURF flow will be improved, and also the maximal number of selected features is kept in order to reduce the computational complexity. Besides, the distribution of optical flow is tuned to be more uniform, which turns out to be important for reducing drift rate of motion estimation. Based on the accurate and uniformly distributed SURF flow, a fundamental matrix estimation combining RANSAC and the eight-point algorithm is applied to compute the relative pose between every consecutive pair of frames. Experiments show that the ego-motion estimation using the optical flow generated by our method is clearly superior compared to other optical flow based techniques in terms of accuracy.