Uncertainty-Aware Selective Obstacle Perception for Micro-UAVs under Visual Degradation

Reliable obstacle perception on micro-UAVs requires not only accurate predictions but also an indication of when those predictions should not be trusted. This paper presents an uncertainty-aware selective perception framework based on a lightweight MobileNetV۳-small multi-task network for obstacle-distance estimation and three-class risk classification. Monte Carlo dropout is used to estimate distance uncertainty and predictive entropy under clean, low-light, blurred, noisy, and partially occluded images. Predictive entropy detects incorrect risk predictions with an AUROC of ۰.۹۰۴, while uncertainty-based rejection improves risk accuracy from ۹۳.۰۴% at full coverage to ۹۸.۶۰% at ۸۰% coverage. The results demonstrate the value of uncertainty-aware selective prediction for safety-related micro-UAV perception.

کلیدواژه ها:

micro-UAV ، uncertainty estimation ، obstacle perception ، selective prediction ، visual degradation