Reading the Fog: Real-Time LiDAR-Based Detection Range Estimation for Autonomous Vehicles in Adverse Weather

Detecting objects with LiDAR in fog has long been treated as a noise-removal problem: identify the fog returns and throw them away. This work takes the opposite view. The same fog returns that other methods discard are read as a direct measurement of how far the sensor can still see. After an off-the-shelf fog-segmentation step, the near-range fog points are used to estimate the atmosphere’s optical thickness, which is then converted into a per-frame maximum detection range (MDR)—a single, human-readable number, in meters, that tells planning and safety modules how far the LiDAR is currently usable. The estimator has no learnable parameter (the physics core—a per-frame Beer–Lambert slope fit, followed by the Koschmieder mapping) and runs in real time during driving. We validate it with real-vehicle experiments in controlled heavy-fog scenarios. Across six fog runs, the estimated MDR recovers the true first-detection distance with a mean absolute error of 1.99 m and R2 = 0.741—a 58% error reduction over a no-feature baseline. To rule out the possibility that the result is an artifact of our own recording, the same estimate is cross-checked against two independent references used by the wider community: the STF transmissometer dataset and the Hahner et al. automotive fog simulator. Our estimate agrees with both, which is its key differentiator from prior LiDAR-fog work that reports performance only on its own data. The result is a real-time, parameter-free, externally corroborated visibility readout that fits inside an existing fog-denoising pipeline and gives downstream autonomy a quantitative answer to the question: “How far can I see right now?”