استفاده از دانش منطقي پس‌زمينه در بهبود تشخيص رويدادهاي جاده‌اي

Autonomous driving systems, as one of the most complex applications of artificial intelligence, require a combination of powerful environmental perception capabilities an‎d precise logical reasoning. While deep neural networks have demonstrated remarkable ability in pattern recognition an‎d visual data processing, they have fundamental limitations in logical reasoning an‎d safety constraint compliance. This research focuses on developing novel methods for integrating symbolic knowledge an‎d logical requirements into the deep learning process to improve model performance an‎d provide more reliable an‎d safer autonomous driving systems. The ROAD-R dataset, as the first comprehensive dataset for road event detection with explicit logical requirements, provides an appropriate foundation for investigating this challenge. This dataset includes 243 logical constraints that model relationships between agents, actions, an‎d locations. Deep neural networks are inherently data-driven, an‎d their optimal performance requires access to a significant volume of labeled data. In practice, providing an‎d accurately annotating such data, especially in the autonomous driving domain, is a time-consuming an‎d expensive process that presents serious practical challenges. In this research, three innovative methods are proposed to address this problem. The first method is based on graph learning of label relationships, which models logical dependencies between different classes using graph convolutional networks. This method provides a framework for understan‎ding visual content an‎d logical constraints by utilizing the YOLOv8 model for object detection an‎d a graph structure for learning relationships. The second method employs score-based diffusion models with neuro-symbolic guidance. This method uses implicit denoising diffusion models an‎d has the capability to generate multimodal label distributions that are simultaneously consistent with visual evidence while respecting logical constraints. The third method is based on selec‎tive unlearning, which offers an efficient solution for more effective fine-tuning of pre-trained models. This method enables model improvement without requiring complete retraining by identifying an‎d selec‎tively removing patterns that lead to logical constraint violations. The results of extensive experiments on the ROAD-R dataset demonstrate that all three proposed methods have performed very well compared to baseline models, an‎d the third method achieved the highest performance, outperforming the best baseline by approximately one point in F1‑Score at intersection over unio‎n of 0.5. This research shows that integrating symbolic knowledge into the deep learning process can improve overall system performance even in scenarios with limited labeled data.

مهران صفاياني , عبدالرضا ميرزايي

محمدعلي خسروي فرد , محمدرضا احمدزاده , عليرضا بصيري