Revisiting the Non-recurrent Traffic Incident Identification Problem for Real-Time Applications Using Theory-Aware Unsupervised Learning

Abstract In this paper, we present a novel approach for real-time detection of non-recurrent traffic patterns in urban roadway networks leveraging advanced machine learning techniques explained by traffic flow theory. The methodology comprises two key components. First, an LSTM-based Autoencoder is employed to extract typical expected traffic patterns from raw traffic data. Second, a clustering technique is utilized to identify non-recurrent congestion, applied on the deviation of the speed and volume measurements from the aforementioned typical patterns. The methodology is implemented using high-resolution multimodal traffic data from the urban road network of Athens, Greece. Findings reveal the presence of four distinct traffic states, three of which represent various types of non-recurrent traffic conditions, characterized by their deviations from typical traffic patterns. The methodology can promptly detect the shift between recurrent and non-recurrent conditions in real time and may have far reaching implications for efficient urban traffic management systems.