Long-Term LiDAR Place Recognition Based on Phase Congruency in Changing Environments

Robust Place recognition is an open challenging problem under various perceptual conditions, e.g., all weather, times-of-day and seasons shifts. Although LiDAR sensors do not suffer from illumination problem of visual appearance, they are dramatically affected by structural changes across seasons. In this paper, a simple yet effective place recognition method is proposed based on phase congruency to achieve long-term robust localization under severe seasonal changes. To improve the efficiency while retaining geometric information of scene layouts in 3D space, a compact cylindrical description is designed to convert 3D point clouds to 2D images, in which the pixel values are computed based on geometric information rather than heights or numbers of points. Then the histogram of orientated phase congruency are encoded as global descriptor for place retrieval. Extensive real experiments on across-season Oxford RobotCar dataset show that the proposed system outperforms baseline methods, verifying its robust performance in challenging scenarios.