Robust Visual Trash Detection for Autonomous Highway Service Area Cleaning Under Severe Weather Conditions

Typhoons generate an unforgiving triad of hurricane-force winds, torrential rain, and dense advection fog that instantly transforms highway service areas (HSAs) into dangerously cluttered landscapes. Rapid litter removal is mandatory to restore traffic throughput and mitigate secondary accidents; however, human cleaning crews are hamstrung by low visibility and slippery surfaces. We present TyphoTrashNet, a perception framework that enables autonomous cleaning vehicles (ACVs) to detect and localise diverse trash objects in real time despite severe meteorological degradation. Our contributions are three-fold: (i) a physics-constrained rain–fog augmentation (RFA) pipeline that synthesises label-preserving images with controllable scattering, noise, and contrast attenuation achieving this with a minimal processing overhead of less than 0.2 ms per image on a standard GPU; (ii) a multi-scale strip spatial attention (MS-SSA) module strategically injected into four stages of YOLOv11-X to reconcile anisotropic context with lightweight computation; and (iii) the 12 k-image TCAO-Typhoon dataset with 98 k bounding boxes spanning 11 litter classes and 27 weather parameters. On this benchmark, TyphoTrashNet lifts mAP50 from 0.052 to 0.415 while maintaining 118 FPS on NVIDIA AGX Orin, outperforming channel (SE) and spatial (CBAM) attention baselines at a lower FLOP overhead. Cross-weather evaluations on RainCityScapes and NightCity demonstrate superior domain generalisation, and comparisons with selected state-of-the-art adverse weather detectors further validate its competitive performance. Field trials on three coastal expressways show a 44 % reduction in clean-up time compared with manual squads. Code and data will be released upon publication.