Lightweight hybrid foreign object detection framework with scene aware and uncertainty fusion mechanism

Foreign object detection (FOD) in autonomous driving environments poses a significant challenge due to the dual nature of anomalous objects: known objects appearing in inappropriate contexts or truly novel objects unseen during training. Existing methods typically address either closed-set semantic reasoning or open-set uncertainty estimation, limiting their ability to handle both types of anomalies effectively. This paper introduces HSAOSFOD (Hybrid Scene-Aware Open-set Foreign Object Detection), a unified lightweight framework that integrates closed-set and open-set paradigms through explicit modeling of a scene-object compatibility matrix and multi-signal open-set detection modules. The key contributions include: (1) a scene-object compatibility module that leverages domain priors to detect contextual mismatches of known objects, (2) a multi-signal fusion module that combines prototype-based novelty detection and uncertainty estimation, and (3) an efficient architectural design utilizing separable self-attention and depth wise convolutions, achieving a model size of only 4.33 million parameters. Extensive experiments on the Cityscapes and RailSem19 datasets for training, and Road Anomaly and Lost and Found datasets for evaluation, demonstrate that HSAOSFOD attains competitive performance with AUROC scores of 0.9506 on Road Anomaly and 0.6158 on Lost and Found, while preserving computational efficiency. Ablation studies confirm that the compatibility module (closed-set) contributes approximately 1.1% and the novelty detection head (open-set) contributes approximately 0.4% to average AUROC, together describing for a combined hybrid contribution of 1.5% over the base decoder alone. HSAOSFOD illustrates the potential of combining explicit domain knowledge with data-driven learning to produce efficient and interpretable hybrid models, delivering particularly strong results on context-sensitive anomalies.