Enhancing vision representations for traffic safety-critical events via supervised contrastive learning.

Area of Science:

• Computer Vision
• Machine Learning
• Road Safety

Background:

• Developing models for traffic safety-critical events (SCEs) like crashes is vital for road safety.
• Current methods struggle to differentiate SCEs from normal driving in video data.

Purpose of the Study:

• To propose a novel supervised contrastive learning (SCL) approach for enhancing traffic video representations.
• To improve the detection of safety-critical events (SCEs) in real-world driving scenarios.

Main Methods:

• Implemented a supervised contrastive learning (SCL) method integrating label information into contrastive loss.
• Utilized a lightweight video encoder optimized for traffic video data.
• Enhanced intra-class cohesion and inter-class separation in the representation space.

Main Results:

• Evaluated on the SHRP2 Naturalistic Driving Study dataset.
• Demonstrated superior performance in representation clustering compared to benchmarks.
• Achieved better results in downstream three-way event classification tasks.

Conclusions:

• The SCL approach yields improved visual representations for traffic video analysis.
• Enhanced SCE detection can significantly benefit advanced driver assistance and automated driving systems.
• This contributes to timely alerts, collision avoidance, and reduced crash risk.