Real-Time and Uniform Feature Filtering for Stereo Images Based on Spatial Indexing

For mobile robots' or vehicles' navigation in dynamic environments, visual odometry is an effective technique to get precise locations at a lower cost. The feature-based method now becomes the typical solution, but it requires heavy computational power in filtering good features and matches. This paper presents an efficient pipeline to detect, filter, and match the features for stereo cameras based on AGAST, BEBLID descriptor, and spatial indexing. We use two quadtrees to manage features and filter good matches by stereo line searching and quadtree querying. The algorithm is evaluated on the KITTI dataset and greatly outperforms a mainstream feature filtering method using an image grid, KNN, and Lowe ratio test in terms of inlier ratio, uniformity, and real-time processing. The improved pipeline is well suited for a full stereo visual odometry or visual SLAM system.