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

The development of models capable of learning visual representations from traffic safety-critical events (SCEs), including crashes and near-crashes, is crucial for advancing road safety.

This study proposes a novel supervised contrastive learning (SCL) approach designed to enhance representations of traffic video data for detecting SCEs. SCL integrates supervised and contrastive learning by embedding label information into the contrastive loss, enhancing intra-class cohesion and inter-class separation in the representation space. Leveraging a lightweight video encoder, SCL optimizes representations specifically for video data, addressing the challenges of distinguishing characteristics between SCEs and normal driving scenarios.

The proposed approach was evaluated using the Second Strategic Highway Research Program Naturalistic Driving Study video dataset, demonstrating superior performance in representation clustering and downstream three-way event classification tasks compared to benchmark approaches.

The improved representations learned through SCL can be directly applied to enhance real-time detection of SCEs in advanced driver assistance systems and automated driving systems. By enabling more reliable identification of SCEs, SCL can support timely driver alerts, proactive collision avoidance maneuvers, and post-incident analysis, ultimately contributing to reduced crash risk and improved road safety.