Feature Matching in the Changed Environments for Visual Localization

Robust feature matching is a fundamental capability for visual SLAM. It remains, however, a challenging task, particularly for changed environments. Some researchers use semantic segmentation to remove potentially dynamic objects which cause changes in the indoor environment. However, removing these objects may reduce the quantity and quality of feature matching. We observed that objects are moved but the room layout does not change. Inspired by this, we proposed to leverage the room layout information for feature matching. Considering current image matching datasets do not have obvious changes caused by dynamic objects and lack layout information, we created a dataset named Changed Indoor 10k (CR10k) to evaluate if we can utilize layout information for image matching in the changed environment. Our dataset contains apparent movements of large objects, and room layout can be extracted from it. We evaluate the performance of existing image matching methods on our dataset and ScanNet dataset. In addition, we propose Layout Constraint Matching (LCM) which is robust to changed environments and the LCM outperforms conventional approaches on the task of pose estimation.