Special Traffic Event Detection: Framework, Dataset Generation, and Deep Neural Network Perspectives.

Area of Science:

• Computer Science
• Artificial Intelligence
• Traffic Engineering

Background:

• Efficient traffic control management relies on the early identification of special traffic events like accidents and road debris.
• Existing systems lack robust capabilities for real-time detection and reporting of diverse highway emergencies from vehicles.
• The need for an integrated in-vehicle system for special traffic event detection and safe driving monitoring is critical.

Purpose of the Study:

• To propose a framework for an in-vehicle module-based system for special traffic event and emergency detection.
• To enhance the efficiency of traffic management on highways through improved detection algorithms.
• To develop and validate deep learning models for identifying critical highway incidents.

Main Methods:

• Development of an in-vehicle module-based framework for real-time traffic event detection.
• Adaptation and application of modified ResNet classification algorithms for special traffic event identification.
• Utilization of datasets containing road debris and vehicle malfunction/accident data from Korean highways.

Main Results:

• Demonstrated feasibility of adapted deep learning algorithms for detecting actual highway emergencies.
• Development of a specialized dataset and detection algorithm tailored for highway incidents.
• Achieved a 9.2% performance improvement compared to object accident detection-based algorithms.

Conclusions:

• The proposed in-vehicle system framework effectively detects special traffic events and enhances safe driving monitoring.
• Modified ResNet classification shows significant potential for improving highway traffic management efficiency.
• The developed algorithm and dataset provide a tailored solution for real-world highway emergency detection.