Stereo Odometry Based on Careful Frame Selection

In this paper, we present a novel stereo visual odometry that provides states-of-the-art results. Our method firstly selects stable features by quad-matching testing and grid-based motion statistics. Instead performing odometry by using past consecutive frames as traditional methods, we select a reference frame by tracing the past frames back as far as possible which shares the stable features with the current frame, and odometry is performed using these two frames. In this way, we skip the frames laying between the two frames, and avoid the accumulation of errors across these frames. Experiments results show that our method outperforms other traditional visual odometry methods on KITTI dataset.