Degradation-Aware Lightweight Vision with Few-Shot Calibration for Micro-UAV Obstacle Perception

Reliable visual perception on micro aerial vehicles is challenging because lightweight onboard models must operate under limited computational resources while remaining robust to variations in illumination, image quality, and scene appearance. This paper presents a degradation-aware lightweight vision framework for joint obstacle-distance estimation and risk classification using MobileNetV۳-small as a shared feature extractor. The network includes two prediction branches for distance regression and three-class risk classification. Its robustness is evaluated under low illumination, Gaussian blur, Gaussian noise, and partial occlusion. The baseline model achieves a distance mean absolute error of ۰.۱۰۴ m and a risk-classification accuracy of ۹۵.۳۳% on clean synthetic images, but its performance degrades substantially under low-light and noisy conditions. A degradation-aware fine-tuning strategy is therefore introduced by applying randomized camera corruptions during training. This reduces distance error by ۸۹.۳۹% under low illumination and ۷۹.۷۷% under Gaussian noise, while restoring risk accuracy to above ۹۵% in both conditions. To reduce the remaining simulation-to-real gap, a lightweight few-shot linear calibration stage is fitted using ۲۴ labeled real images and evaluated on ۱۲ unseen Crazyflie-camera images across five balanced data splits. The calibrated model achieves a mean distance error of ۰.۳۰۶±۰.۰۷۰ m and a risk accuracy of ۶۸.۳۳±۱۰.۸۷%, compared with ۰.۷۵۷±۰.۰۱۷ m and ۴۶.۶۷±۷.۴۵% for the uncalibrated degradation-aware model. These results indicate that degradation-aware training combined with low-cost output calibration can improve robust obstacle perception without modifying the lightweight network architecture.

کلیدواژه ها:

micro aerial vehicles ، lightweight vision ، obstacle perception ، degradation-aware learning ، few-shot calibration ، simulation-to-real transfer