Knowledge Distillation for Efficient and Robust On-Board Motorcycle Perception in Low-Light Conditions

Adaptive matrix LED headlights are a promising safety feature for motorcycles, dynamically adjusting illumination to avoid blinding other drivers while preserving long-range visibility. Enabling this behaviour requires a perception system able to detect surrounding vehicles and estimate their metric depth in real time. This is especially challenging in the motorcycle domain, where perception must run in real time on constrained onboard hardware, rely on a single monocular camera, and cope with the motion blur induced by the vehicle's speed and dynamics. These challenges become even more pronounced at night, where reduced visibility, strong illumination variations caused by vehicle headlights, and low-contrast scenes significantly degrade the reliability of vision-based perception. Moreover, no annotated data is available for this task. This thesis addresses these constraints with a teacher-student pseudo-labeling pipeline that extracts supervisory signals for both detection and metric depth estimation from unannotated nighttime footage, paired with a lightweight multi-task model designed for real-time embedded deployment. The resulting system is evaluated both on the held-out test split of the pseudo-labeled dataset and against a pinhole-based ground-truth set built for this work, on which the distilled student matches the accuracy of the zero-shot teacher it was distilled from, at a fraction of its computational cost.