IncidentNet: Traffic Incident Detection, Localization and Severity Estimation with Sparse Sensing

TL;DR

This work tackles rapid, accurate urban traffic incident detection under sparse sensor coverage by generating realistic microscopic datasets from macroscopic data and presenting IncidentNet, a three-task deep-learning pipeline. A synthetic pipeline using Tempe data models microscopic flows via a two-frequency sinusoidal representation and Levenberg–Marquardt optimization, simulates incidents, and builds a sparse-sensing dataset with SUMO/Traci. IncidentNet leverages a TabNet-based, cascaded architecture to detect incidents, localize their roads, and estimate severity, performing robustly under sparse sensor conditions and delivering real-time inference. The approach achieves high urban DR (≈98%), low FAR (≈6%), and competitive MTTD, with highway applicability demonstrated, underscoring practical potential for sparse surveillance deployments and emergency response optimization.

Abstract

Prior art in traffic incident detection relies on high sensor coverage and is primarily based on decision-tree and random forest models that have limited representation capacity and, as a result, cannot detect incidents with high accuracy. This paper presents IncidentNet - a novel approach for classifying, localizing, and estimating the severity of traffic incidents using deep learning models trained on data captured from sparsely placed sensors in urban environments. Our model works on microscopic traffic data that can be collected using cameras installed at traffic intersections. Due to the unavailability of datasets that provide microscopic traffic details and traffic incident details simultaneously, we also present a methodology to generate a synthetic microscopic traffic dataset that matches given macroscopic traffic data. IncidentNet achieves a traffic incident detection rate of 98%, with false alarm rates of less than 7% in 197 seconds on average in urban environments with cameras on less than 20% of the traffic intersections.

Figures (6)

• Figure 1: The plot of the vehicle counts for a 24-hour period from the Department of Transportation of Tempe for the 12 roads between the placed sensors of interest from the selected Tempe region shown in Fig. \ref{['fig: tempe map with real-world sensor placement']}.
• Figure 2: Representation of averaged ground-truth vehicle counts and generated traffic flow model. To ensure variance in generated vehicle counts, we add a small deviation $alpha$.
• Figure 3: Shows A Tempe, AZ region selected as the test area for our implementations. All the plotted points indicate the locations where cameras can be deployed for simulation. However, the deployed locations are highlighted in red to make the deployment of cameras similar to the real world.
• Figure 4: The block diagram depicts IncidentNet's architecture. The raw data from the simulator is transformed into processed data. For training, all data points are used for the incident detection model, and data points with positive incident labels are used for incident localization and severity estimation models. During the prediction phase, localization and severity estimation models depend on the incident detection model's prediction.
• Figure 5: The KS Statistic and the p-value obtained from the KS test for the four days of data made available by Tempe are shown. The p-value threshold is indicated as the red line.