Hierarchical Transformer-Based Red-Light Running Prediction Model for Two-Wheelers with Multitask Learning

The red-light running behavior of two-wheelers presents significant safety concerns at urban signalized intersections. Anticipating their intentions and trajectories is crucial, as early detection can substantially mitigate potential risks. However, predicting such behavior is challenging due to the interactions among red-light runners, neighboring traffic agents, the local and global traffic environment. To this end, we propose a Hierarchical Transformer-based red-light running prediction model for two-wheelers in connected environments. Our model features a three-level hierarchical encoder, which effectively integrates multi-source input data. A Social-Aware Motion Encoder, Scene Information Encoder and Environment Encoder are first proposed to extract the first-level input features. Subsequently, we propose a Hierarchical Hybrid Encoder and utilize crossattention modules to capture the interdependencies of the extracted features at the second and the third levels, respectively. To facilitate simultaneous prediction of red-light running intention and trajectory, we employ multitask learning and design an anchor-based structure to allow flows of information in different decoders. The proposed model is evaluated using real-world data collected from a signalized intersection over a period of 7 days. Our model demonstrates state-of-the-art performance across various evaluation metrics, validating its efficacy in predicting future red-light running behaviors in dynamic traffic scenarios. Code is available at https://github.com/llllys/RLR-prediction.