ATM-SLAM: Adaptive Temporal Memory Visual SLAM in Severe Rainy Conditions
TL;DR
This work proposes ATM-SLAM, a novel learning-based vSLAM framework that overcomes the intense visual interference under severe rainy conditions and enhances system robustness and collected the Severe Rainy SLAM Dataset (SR-SLAM), a large-scale real-world dataset, and plan to release it as a dedicated benchmark for evaluating vSLAM robustness.
Abstract
Visual Simultaneous Localization and Mapping (vSLAM) plays a crucial role in autonomous vehicles and robotics by enabling real-time localization and mapping in unknown environments. However, as current vSLAM systems are primarily designed for ideal weather conditions they often experience significant accuracy degradation and sudden operational breakdowns in severe rainy conditions (> 10 mm/h), posing serious safety hazards. To fill the gap, we propose ATM-SLAM, a novel learning-based vSLAM framework that overcomes the intense visual interference under severe rainy conditions and enhances system robustness. ATM-SLAM incorporates two key modules: (i) an Adaptive Feature Encoder (AFE), which adaptively prioritizes rain-robust features across consecutive frames, (ii) a Temporal Memory Integration (TMI) module, which integrates memorized correlation information to recover and stabilize localization under distorted frames. We extensively evaluate ATM-SLAM on a consumer vehicle, a robotic platform, and widely used public datasets across diverse scenarios. The evaluation results demonstrate that ATM-SLAM can achieve accurate and reliable localization and mapping in heavy rain. Moreover, to facilitate future research on vSLAM under severe rainy conditions, we collected the Severe Rainy SLAM Dataset (SR-SLAM), a large-scale real-world dataset, and plan to release it as a dedicated benchmark for evaluating vSLAM robustness.