CENH: Co-Enhancement of Edges and Nodes in Heterogeneous Graph Attention for Pedestrian Crossing Intention Recognition

Accurately considering the interactions between the target pedestrian and surrounding scene elements is crucial. It helps capture behavioral changes and understand the mutual influence between the target and its environment. This, in turn, plays a key role in enhancing intention recognition. However, with the ego-vehicle’s front-facing camera, challenges arise from the difficulty in quantifying the distance between objects and the variation in the observed size of objects with distance, which limits the ability to fully account for interactions in existing studies. To address these challenges, we propose the Co-Enhancement of Edges and Nodes in Heterogeneous Graph Attention (CENH) method for pedestrian crossing intention recognition. This approach effectively incorporates the impact of interactions between the target pedestrian and scene elements, overcoming issues like the unification of interaction ranges, quantification of interaction strength, and modeling of interactions from the ego-vehicle’s perspective. Specifically, we design a method to define the interaction range based on the distance between the target and the ego-vehicle, adaptively selecting relevant scene elements. We employ a directed heterogeneous graph to represent the dynamic interactions between the target and these elements, enhancing edge and node attributes using graph attention. Additionally, a local sparse attention module is introduced to extract key local features of the target, which are then used to predict pedestrian crossing intention. Experiments on the PIE and JAAD datasets demonstrate that CENH consistently outperforms existing methods across different prediction horizons. The proposed method effectively captures and exploits the interactions between target pedestrians and surrounding scene elements, while maintaining strong stability, adaptability, and robustness in complex traffic scenarios.