Intention Transformer for driving intention and trajectory prediction in dilemma zones of signalized intersection

• Extract complex features for driving behavior prediction under yellow dilemma at holographic intersections. • Propose a real-time prediction framework for driving intention and trajectory in yellow dilemma zone. • Identify complex driving decision processes with intentional changes under single-vehicle and multi-vehicle scenarios. • Conduct a case study at a real-world intersection in Ningbo, China; • Outperform state-of-the-art models in driving intention and trajectory prediction tasks. Red-light-running is a primary cause of traffic accidents at signalized intersections. Analyzing and predicting driving intention and trajectory under the yellow dilemma is critical in preventing traffic accidents. With the development of intelligent transportation technology, an increasing number of cities have deployed holographic intersections that utilize radar and video sensors to collect all-weather, high-resolution, and all-around traffic data. This paper proposes a real-time multi-step framework, Intention Transformer, which integrates holographic intersection vehicle trajectory information, high-definition map information, and signal timing information to predict driving intention and trajectory under the yellow dilemma. Our proposed framework is based on the Transformer architecture, which comprises an intention feature extraction module, a trajectory feature extraction module, and an interaction module that combines intention and trajectory features. The intention and trajectory loss functions are jointly optimized, and the intention probability and future trajectory are predicted at each step. We conducted a case study at a real-world intersection in Ningbo, China. The results show that our proposed framework achieves the best performance compared with state-of-the-art models in driving intention and trajectory prediction tasks. The ablation experiment demonstrates that driving intention and trajectory features can enhance each other’s predictive performance. Analyzing driving behavior enables us to identify complex driving decision processes with intentional changes both under single-vehicle and multi-vehicle scenarios. This paper proposes a novel approach to driving behavior prediction in dilemma zones at holographic signalized intersections, which captures and predicts complex driving behavior features in real time, and lays a firm groundwork for fine-grained urban road traffic management.