Enhancing Traffic Incident Detection Through ADASYN-Attention Fusion: A Comparative Study with RITIS Data

Traffic incidents are a leading contributor to non-recurring congestion and secondary crashes. Each year congestion and crashes together cost the United States over 1 trillion dollars. Once traffic queues are formed, it is difficult to dissipate them and return traffic to normal operations. Therefore, real-time and accurate incident detection plays a critical role in Traffic Incident Management (TIM). This research focuses on highway traffic incident detection. It divides a highway network into short segments and correlates temporal and spatial data from adjacent segments for detecting incidents. Due to incidents being relatively rare compared to normal traffic patterns, we propose a method that combines oversampling with the attention mechanism and use an ablation study to prove its effectiveness in improving supervised incident detection.