Abstract
1 min readTraffic congestion, accidents, and unpredictable driver behaviour remain significant challenges in urban transportation systems. Traditional traffic simulation models often fail to adapt to dynamic environments and lack accuracy handling edge-case scenarios. To address these limitations, hybrid TrafficAI, an innovative Generative AI-based framework that integrates advanced modules for traffic simulation, behaviour modelling and anomaly detection. The framework incorporates several key components. First, an Adaptive Multi-Modal Fusion Engine (AMFE) seamlessly integrates video, LiDAR, and textual data. This is achieved through dynamic feature alignment layers and context-aware gating mechanisms. Second, an Edge-Case Generative Module (ECGM) augments synthetic edge-case scenarios. Third, a Temporal-Spatial Attention Network (TSAN) captures short-term and long-term traffic dependencies. Finally, large language model-driven semantic reasoning modules extract contextual insights from unstructured textual data, such as traffic reports and incident logs. The framework employs a hybrid dual-stage optimization process, combining unsupervised generative pre-training with fine-tuned supervised calibration to ensure efficient convergence and reduced latency. By fusing multi-modal data, enhancing anomaly robustness with synthetic edge-case scenarios, and interpreting contextual semantics with LLM, hybrid TrafficAI achieves precise anomaly detection, trajectory prediction and adaptive decision-making. Experimental evaluations demonstrate significant performance improvements, including 91.45% accuracy, 93.45% mean Average Precision (mAP) for vehicle detection and a 0.910 Normalized Risk Score (NRS) for anomaly detection, consistently outperforming state-of-the-art benchmarks across latency, precision, recall and stability metrics. This framework sets a new benchmark for intelligent transportation systems (ITS) and real-time traffic management.
Discussion(0)
No comments yet. Be the first to comment.