Learning 3D Segment Descriptors for Place Recognition
This work addresses place recognition for localization and navigation in autonomous driving, but it is incremental as it builds on existing segment extraction and matching approaches.
The paper tackles place recognition in LiDAR-based 3D point cloud maps by proposing a learning-based method using CNNs to generate segment descriptors, achieving higher recall accuracy than hand-crafted descriptors in urban driving scenarios.
In the absence of global positioning information, place recognition is a key capability for enabling localization, mapping and navigation in any environment. Most place recognition methods rely on images, point clouds, or a combination of both. In this work we leverage a segment extraction and matching approach to achieve place recognition in Light Detection and Ranging (LiDAR) based 3D point cloud maps. One challenge related to this approach is the recognition of segments despite changes in point of view or occlusion. We propose using a learning based method in order to reach a higher recall accuracy then previously proposed methods. Using Convolutional Neural Networks (CNNs), which are state-of-the-art classifiers, we propose a new approach to segment recognition based on learned descriptors. In this paper we compare the effectiveness of three different structures and training methods for CNNs. We demonstrate through several experiments on real-world data collected in an urban driving scenario that the proposed learning based methods outperform hand-crafted descriptors.