For Autonomous Vehicles (AVs), recognizing traffic lights and signs is critical for safety because perception errors directly affect navigation decisions. Real-world disturbances such as glare, rain, dirt, and graffiti, as well as digital adversarial attacks, can lead to dangerous misclassifications. Current research lacks (i) temporal continuity (stable detection across consecutive frames to prevent flickering misclassifications), (ii) multi-field-of-view (FoV) sensing, and (iii) integrated defenses against both digital and natural degradation. This paper presents two principal contributions: (1) a three-layer defense framework integrating feature squeezing, inference-time temperature scaling (softmax τ = 3 without distillation training), and entropy-based anomaly detection with sequence-level temporal voting; (2) a 500 sequence dual-FoV benchmark (30k base frames, 150k with perturbations) from aiMotive, Waymo, Udacity, and Texas sources across four operational design domains. The unified defense stack achieves 79.8% mAP on a 100-sequence test set (6k base frames, 30k with perturbations), reducing attack success rate from 37.4% to 18.2% (51% reduction) and high-risk misclassifications by 32%. Cross-FoV validation and temporal voting enhance stability under lighting changes (+3.5% mAP) and occlusions (+2.7% mAP). Defense improvements (+9.5–9.6% mAP) remain consistent across native 3D (aiMotive, Waymo) and projected 2D (Udacity, Texas) annotations. Preliminary recapture experiments (n = 15 scenarios) show 2.5% synthetic–physical ASR gap (p = 0.18), though larger validation is needed. Code, models, and dataset reconstruction tools are publicly available.
Joshi et al. (Mon,) studied this question.