Self-localization on a 3D map by fusing global and local features from a monocular camera
Satoshi Kikuchi, Masaya Kato, Tsuyoshi Tasaki
TL;DR
The paper addresses monocular camera self-localization on a 3D map in the presence of dynamic obstacles by fusing CNN-derived local features with Vision Transformer-based global features. It introduces a three-block network that separately extracts global and local features and then regresses pose from their fusion, demonstrating improvements over the state-of-the-art LHMap-loc on public datasets (KITTI and nuScenes) and in CG/dynamic scenarios. Key findings include a 20.1% reduction in self-localization error on nuScenes and an average robot localization accuracy of 7.51 cm, highlighting robustness to dynamic objects and map discrepancies. The work suggests that integrating global context via ViT with local CNN features enhances camera-based localization, with future work aiming to further stabilize estimates using particle filters.
Abstract
Self-localization on a 3D map by using an inexpensive monocular camera is required to realize autonomous driving. Self-localization based on a camera often uses a convolutional neural network (CNN) that can extract local features that are calculated by nearby pixels. However, when dynamic obstacles, such as people, are present, CNN does not work well. This study proposes a new method combining CNN with Vision Transformer, which excels at extracting global features that show the relationship of patches on whole image. Experimental results showed that, compared to the state-of-the-art method (SOTA), the accuracy improvement rate in a CG dataset with dynamic obstacles is 1.5 times higher than that without dynamic obstacles. Moreover, the self-localization error of our method is 20.1% smaller than that of SOTA on public datasets. Additionally, our robot using our method can localize itself with 7.51cm error on average, which is more accurate than SOTA.