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

Background

Traffic accidents are increasingly common in recent years, in part caused by traffic incidents, road maintenance, and emergency scenarios. Recent studies have shown that even a five-minute delay in emergency response to traffic accidents result in a 46% increase in fatality rates, while response times under seven minutes reduce fatality rates by 58% in urban and rural areas. Current technologies for traffic incident detection revolve around high sensor coverage, and are primarily based on decision-tree and random forest models that have limited representation capacity and cannot detect incidents with high accuracy.

Invention Description

Researchers at Arizona State University have developed IncidentNet, which is a new deep learning approach to accurately detect, localize, and assess the severity of traffic incidents using sparse sensor data. IncidentNet generates synthetic microscopic traffic data to overcome the scarcity of suitable datasets, achieving fast detection rates and low false alarm rates, which significantly enhances urban traffic management capabilities. IncidentNet improves the accuracy over existing decision-tree and random forest models in traffic incident detection.

Potential Applications:

• Urban traffic management and control systems
• Emergency response optimization
• Intelligent transportation systems (ITS)
• Research and development in traffic flow and incident analysis
• Smart city infrastructure planning and development

Benefits and Advantages:

• High detection rate – achieves nearly 98% incident detection rate
• Low false alarm rates – overcomes the scarcity of suitable datasets, achieving false alarm rates of less than 7%
• Effective – functions in environments with cameras covering less than 20% of intersections
• Improved model training – capable of generating synthetic traffic data
• Rapid detection – supports detection within 197 seconds on average