Multi-stage classification of abnormal traffic events using a multi-head + LSTM

Traffic congestion, anomalies and incident significantly impact urban transportation efficiency and road safety. Accurate detection and classification of such events are crucial for effective traffic managements, emergency response and infrastructure planning. Traditional approaches based statistical and conventional machine learning models often struggle to generalize across the dynamic and complex traffic patterns which evolves around the time. To address these limitations, we proposed a multi-head + LSTM model in a multistage classification framework. This proposed framework systematically detects anomalies using isolation forest, classifies the congestion into low, medium, high using K-means clustering and determine whether an incident caused an anomaly using a spatial threshold-based approach (1.5 km). the model is trained on 15 days of PeMS traffic data integrated with weather information to enhanced predictive accuracy. Through hierarchical classification the proposed model captures temporal dependencies, integrates contextual weather information and ensures robust anomaly detection, congestion classification and incident identification. Experimental results demonstrates that the multi-Head model significantly outperforms existing methods achieving higher precision, recall, f1-score and ROC-AUC across all classification stages. The results highlight the potential of deep learning-based traffic analysis for intelligent transportation system (ITS) enabling data-driven decision making for urban traffic management.