A preprint posted on arXiv on 9 October by Michigan Technological University researchers aims to detect the wheel tracks left by earlier vehicles on snow-covered, visually sparse roads, rather than mark a broad drivable area. The method combines visible RGB images with long-wave infrared thermal imagery. Training labels are generated automatically by projecting the future wheel path from GNSS records onto images. The team curated a 951-image sequence from the WADS winter-conditions dataset and expanded the training set to 7,608 images through augmentation.

In the paper’s results table, the RGB-and-thermal fusion model recorded an AUROC of 0.8744, compared with 0.6025 for the RGB-only model; false-negative rates were 0.2399 and 0.7927, respectively. The authors report that thermal input helps recover tracks missed by RGB in low-contrast conditions. These findings are limited to one curated winter-driving sequence and a research prototype; the paper does not demonstrate closed-loop driving in a real vehicle or safe operation across all weather conditions. The work is a preprint that has not yet undergone peer review.