A time-embedded attention-based transformer for crash likelihood prediction at intersections using connected vehicle data

The real-time crash likelihood prediction model is an essential component of the proactive traffic safety management system. Over the years, numerous studies have attempted to construct a crash likelihood prediction model in order to enhance traffic safety, but mostly on freeways. In the majority of the existing studies, researchers have primarily used a deep learning-based framework to identify crash potential. Lately, Transformers have emerged as a potential deep neural network that fundamentally operates through attention-based mechanisms. Transformers exhibit distinct functional benefits over established deep learning models like Recurrent Neural Networks (RNNs), Long Short-Term Memory networks (LSTMs), and Convolutional Neural Networks (CNNs). First, they employ attention mechanisms to accurately weigh the significance of different parts of input data, a dynamic functionality that is not available in RNNs, LSTMs, and CNNs. Second, they are well-equipped to handle dependencies over long-range data sequences, a feat RNNs typically struggle with. Lastly, unlike RNNs, LSTMs, and CNNs, which process data in sequence, Transformers can parallelly process data elements during training and inference, thereby enhancing their efficiency. Apprehending the immense possibility of Transformers, this paper proposes inTersection-Transformer (inTformer), a time-embedded attention-based Transformer model that can effectively predict intersection crash likelihood in real-time. The inTformer is basically a binary prediction model that predicts the occurrence or non-occurrence of crashes at intersections in the near future (i.e., next 15 min). The proposed model was developed by employing traffic data extracted from connected vehicles. Acknowledging the complex traffic operation mechanism at intersection, this study developed zone-specific models by dividing the intersection region into two distinct zones: within-intersection and approach zones, each representing the intricate flow of traffic unique to the type of intersection (i.e., three-legged and four-legged intersections). In the ‘within-intersection’ zone, the inTformer models attained a sensitivity of up to 73%, while in the ‘approach’ zone, the sensitivity peaked at 74%. Moreover, benchmarking the optimal zone-specific inTformer models against earlier studies on crash likelihood prediction at intersections and several established deep learning models trained on the same connected vehicle dataset confirmed the superiority of the proposed inTformer. Further, to quantify the impact of features on crash likelihood at intersections, the SHAP (SHapley Additive exPlanations) method was applied on the best performing inTformer models. The most critical predictors were average and maximum approach speeds, average and maximum control delays, average and maximum travel times, split failure percentage and count, and percent arrival on green.