LiteFT-PR: Lightweight and Fault-Tolerant LiDAR-Camera Fusion Network for Robust Place Recognition via Model Distillation

Abstract

Place recognition (PR) is a key component of simultaneous localization and mapping (SLAM) in autonomous vehicles and robotics. By efficiently matching descriptors generated from the current scene with a prebuilt reference database, existing PR methods enable accurate vehicle re-localization. However, achieving robust and efficient PR remains challenging, particularly in scenarios with sensor corruptions, such as camera overexposure and LiDAR beam missing, which degrade the performance of traditional PR systems. To address this issue, we propose LiteFT-PR, a fault-tolerant and lightweight LiDAR-camera fusion PR network. To balance accuracy and efficiency under extreme conditions, we adopt a knowledge distillation framework for PR tasks: the teacher network learns robust representations under corrupted sensing, while the student distills and inherits this knowledge for lightweight deployment. In addition, we design efficient multimodal fusion mechanisms for both networks, enabling robust place recognition under various sensor corruption scenarios while significantly reducing computational overhead. To meet real-world deployment needs, we build a dataset covering multiple types of sensor failures and conduct extensive field experiments. Experimental results demonstrate that LiteFT-PR, with only 7.24M of parameters, outperforms state-of-the-art PR methods in both accuracy and efficiency.