Learning salient representation of crashes and near-crashes using supervised contrastive variational autoencoder.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Traffic Safety

Background:

• Learning salient representations is vital for analyzing safety-critical events (SCEs) like crashes.
• Existing models may not effectively capture the unique characteristics of SCEs.

Purpose of the Study:

• To propose a novel deep learning model, the supervised contrastive variational autoencoder (scVAE), for learning salient representations of SCEs.
• To enhance the discriminative power of latent variables for improved clustering and analysis of driving data.

Main Methods:

• Developed a supervised contrastive variational autoencoder (scVAE) integrating supervised contrastive learning with VAEs.
• Utilized two distinct encoders to promote discriminative salient latent variables.
• Applied the scVAE to kinematic datasets from the SHRP 2 Naturalistic Driving Study.

Main Results:

• The scVAE demonstrated effectiveness in learning salient representations superior to alternative models.
• Clear clustering patterns were observed in the learned representation space.
• Facilitated downstream tasks including sample generation, denoising, and prediction.

Conclusions:

• The scVAE offers a promising approach for learning enhanced representations for traffic safety applications.
• The combination of contrastive and supervised learning is scalable to other frameworks and data types.
• The scVAE contributes to improved driving scenario generation and SCE detection.