BTC: A Binary and Triangle Combined Descriptor for 3-D Place Recognition

Chongjian Yuan,Jiarong Lin,Zheng Liu,Hairuo Wei,Xiaoping Hong,Fu Zhang

DOI: https://doi.org/10.1109/tro.2024.3353076

IF: 7.8

2024-02-10

IEEE Transactions on Robotics

Abstract:Accurate and robust place recognition is essential for robot navigation, yet achieving full pose invariance and high performance across diverse scenes remains challenging. In this work, we propose a novel global and local combined descriptor named binary triangle combined (BTC) descriptor. We first extract the keypoints of a point cloud by projecting the points to planes extracted therein. Any three keypoints form a unique triangle, with the lengths of its sides constituting a triangle descriptor that captures the global appearance of the point cloud. Thanks to the distinct shape of a triangle given three side lengths, the similarity between two triangles and their vertices (i.e., keypoints) correspondence can be naturally determined from the side lengths of the triangle descriptors. The matched triangle pairs evaluate the appearance similarity between two point clouds, while the vertices' correspondence enables accurate estimation of their relative pose; both are crucial for the place recognition task. To enhance the accuracy of triangle matching, BTC introduces a binary descriptor, which describes the point distribution neighboring each keypoint. The local geometry information encoded by the binary descriptor augments descriptiveness and discriminativeness to the triangle descriptor. Collectively, the two descriptors achieve both global and local descriptions of the environment with high accuracy, efficiency, and robustness. We extensively compare the proposed BTC descriptor against state-of-the-art methods (e.g., Scan Context, LCD-Net) on a wide range of datasets collected using different types of LiDAR sensors (spinning LiDARs and nonrepetitive scanning LiDARs) in various environments (urban, campus, forest, park, and mountain). The quantitative results demonstrate that BTC exhibits greater adaptability and significant improvement in precision compared to its counterparts, especially in challenging cases with large viewpoint variations (e.g., reverse direction, large translation, and/or rotation).

robotics