Milo Henry Lovelace Knowles

abstract:We conduct a comprehensive performance comparison of descriptor-based data association methods for object-level SLAM. Based on our findings, we design a visual bag-of-words descriptor from Oriented FAST and Rotated BRIEF descriptors extracted from object bounding boxes. To disambiguate between perceptually similar objects and improve the efficiency of our algorithm, we incorporate a heuristic for filtering associations based on geometric information from the robot pose and object map. We show that our bag-of-words approach with geometric filtering outperforms the precision, recall, and runtime of baseline descriptor matching methods on 21 challenging driving sequences from the KITTI Tracking Dataset.