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

Models capable of learning representations that are salient in safety–critical events (SCEs; including crashes and near-crashes) are crucial for road safety. This study proposes a novel deep learning model, the supervised contrastive variational autoencoder (scVAE), that incorporates supervised contrastive learning methods into the variational autoencoder (VAE) framework. By leveraging two distinct encoders, the scVAE encourages the salient latent variables to be discriminative, capturing the unique representations of SCEs while being regulated by the response variable to focus on the most relevant representations for accurate clustering. Through application on the Second Strategic Highway Research Program (SHRP 2) Naturalistic Driving Study kinematic datasets, we demonstrated the effectiveness of the scVAE in learning salient representations that enable improved clustering compared to alternative models. Quantitative analysis revealed clear clustering patterns in the learned salient representation space, facilitating downstream tasks such as generating samples, denoising, and prediction. The proposed approach of combining contrastive and supervised learning can be scaled to other model frameworks and data modalities, offering a promising direction for learning-enhanced representations that cater to tasks of interest. The study findings highlight the contributions of scVAE to traffic safety, offering enhanced capabilities for driving scenario generation and SCE detection.