CrossFormerGAN: enhancing pedestrian trajectory prediction in autonomous driving through advanced attention...

Pedestrian trajectory prediction is essential for autonomous driving systems, aiming to foresee pedestrian movements and improve safety by anticipating their future positions and paths. Traditional methods often fail to capture the full complexity of pedestrian behavior due to their limited ability to account for subtle gestures, environmental factors, and social interactions, which critically affect movement patterns. This lead to the system making incorrect predictions, potentially leading to unsafe driving decisions. To address this, we propose CrossFormerGenerative Adversarial Network (CrossFormerGAN), a model designed to enhance real-time pedestrian trajectory prediction by introducing Gesture-Spatial Interactive Attention and an Adaptive Neighborhood-based CrossFormer transformer within the generator. Gesture-Spatial Interactive Attention combines Slot Attention and Regional Attention to accurately capture subtle movements and relevant context, ensuring the system thoroughly understands pedestrian intentions. Without this mechanism, important cues like a pedestrian's decision to cross the road could be missed, leading to less reliable predictions. The Adaptive Neighborhood-based CrossFormer transformer integrates an Adaptive Neighborhood Bias Module, which captures sudden changes in movement as well as ongoing movement while considering critical environmental factors like road layouts and traffic signals. This ensures that the model adapts to varying conditions in real-time, preventing potential misjudgements. The generator leverages these insights to encode trajectories of pedestrian into feature tensors and decode them into detailed trajectory predictions. Meanwhile, the discriminator uses Grouped Query Attention instead of multi-head attention to enhance its ability to recognize complex patterns by focusing on different aspects of the data, ensuring that the generated trajectories closely mimic real ones. This comprehensive setup allows the CrossFormerGAN to produce highly accurate and reliable pedestrian trajectory predictions, significantly reducing the risk of accidents. Simulation results confirm that our model outperforms existing methods, achieving the best performance with an ADE of 0.10 on the UCY dataset (ZARA 2 scenario) and an FDE of 0.18 on the ETH dataset (HOTEL scenario) when compared to other scenarios. The proposed method improves the general safety and dependability of autonomous vehicles in dynamic, real-world driving situations in addition to improving forecast accuracy.