13 July 2026 in SSRN Electronic Journal
Low-visibility conditions such as smoke, haze, and dense fog severely degrade the perception and localization capabilities of autonomous systems, yet most existing multi-sensor SLAM methods rely on fixed-weight or hard-switching fusion strategies that cannot adapt to dynamically varying visibility. This manuscript studies a visibility-adaptive multi-sensor fusion SLAM framework that couples a 4D millimeter-wave radar with a LiDAR and a monocular camera. A 4D radar front-end odometry is developed that jointly exploits Doppler velocity, spatial geometry, and Radar Cross Section (RCS) through an improved RCS-weighted correlative scan matching scheme. A lightweight Multi-Scale Spatial-Channel Attention Network (MS-SCANet) is constructed to estimate a continuous environmental visibility score from camera images in real time, using depthwise separable convolutions for multi-scale feature extraction and a spatial-channel dual-attention module for robust degradation representation. The estimated visibility then drives a Cross-Modal Attention Fusion (CMAF) module, which generates sensor-specific weights and embeds them into an Extended Kalman Filter to fuse LiDAR and radar observations for final state estimation. The method is evaluated on a self-collected wheeled-robot dataset covering a dim subterranean tunnel, an outdoor road with patchy smoke, and a smoke-filled indoor corridor, and is further validated on public 4D-radar datasets. Quantitative comparisons against single-sensor and conventional fusion baselines, together with ablation studies, visibility-estimation accuracy analysis, weight-factor analysis, and per-module runtime profiling, show that the proposed method keeps the absolute trajectory error below 0.2m in smoke-filled scenes while running at about 80ms per frame.