Enhancing Pedestrian Trajectory Prediction Through Multi-Stream Complementary Attention Gated Network

Accurately predicting future trajectory of pedestria-ns is crucial for the performance of autonomous systems. However, it is difficult to accurately capture pedestrian behavioral patterns due to the randomness and diversity of pedestrian movements and the complexity of pedestrian interactions, which poses a challenge for prediction. Therefore, this paper proposes a multi- stream complementary attention gated network (MSCAGN) to learn the spatio-temporal features of pedestrian trajectory. Specifically, we use three branches to learn temporal features, spatial interaction features, and spatio-temporal correlation features separately. Meanwhile, complementary attention mechanisms are employed on each branch to capture important features and those that have been overlooked, thereby improving the fitting ability of the model. Finally, Features from the three branches are adaptively fused through gated network. We evaluated the model on the ETH and UCY public datasets, and the experimental results showed that the proposed model achieved state-of-the-art performance in pedestrian trajectory prediction tasks.