3D point cloud-based place recognition: a survey

Abstract

Place recognition is a fundamental topic in computer vision and robotics. It plays a crucial role in simultaneous localization and mapping (SLAM) systems to retrieve scenes from maps and identify previously visited places to correct cumulative errors. Place recognition has long been performed with images, and multiple survey papers exist that analyze image-based methods. Recently, 3D point cloud-based place recognition (3D-PCPR) has become popular due to the widespread use of LiDAR scanners in autonomous driving research. However, there is a lack of survey paper that discusses 3D-PCPR methods. To bridge the gap, we present a comprehensive survey of recent progress in 3D-PCPR. Our survey covers over 180 related works, discussing their strengths and weaknesses, and identifying open problems within this domain. We categorize mainstream approaches into feature-based, projection-based, segment-based, and multimodal-based methods and present an overview of typical datasets, evaluation metrics, performance comparisons, and applications in this field. Finally, we highlight some promising research directions for future exploration in this domain.