Robustness of visual or visual-inertial SLAM systems in extreme environments

Abstract

To address the vulnerability of visual or visual-inertial SLAM systems in environments with drastic illumination variations, sparse textures, and high-dynamic disturbances, this paper proposes a robust multimodal complementary perception SLAM framework. The methodology consists of three core strategies: 1. Thermal infrared modality integration is introduced to compensate for the perception blind spots of visual sensors under extreme lighting conditions. 2. Multi-type feature fusion, incorporating points, lines, and edges, is employed to strengthen geometric constraints in texture-poor environments. 3. A dual-filtering mechanism combining semantic priors with geometric verification is designed to suppress dynamic interference. Furthermore, an adaptive degradation decision-making mechanism based on environmental quality assessment is proposed, enabling smooth mode switching and progressive functional recovery under extreme conditions. Experimental results on the self-collected “Triple-Coupling Extreme Environment” dataset demonstrate that the proposed method achieves an Absolute Trajectory Error (ATE RMSE) of 1.24m, representing an improvement of over 50% compared to state-of-the-art algorithms. Additionally, the system maintains a real-time processing speed of approximately 20.7 fps on an embedded platform.