Detecting vulnerable road users (VRUs) such as pedestrians and cyclists in low light remains a challenge for camera-based safety systems. The most dangerous hours for collisions are also the darkest, but progress has been held back by a lack of large-scale training datasets. A new preprint on arXiv proposes an automatic labeling pipeline that uses stereo long-wave infrared (LWIR) and RGB cameras to generate ground-truth annotations without manual effort.

By combining thermal and visible-spectrum imagery, the method exploits the complementary strengths of both modalities: LWIR captures heat signatures in darkness, while RGB provides texture and color. The stereo setup allows depth information to be used for aligning and validating labels. The authors argue this can reduce the cost of dataset creation and enable more robust VRU detection in low-light conditions.

Because the abstract is brief, details about the labeling algorithm and evaluation results are not yet available. The work addresses a real bottleneck in the field, however, and points toward automated data generation as a way forward.