Graph Convolutional Networks with Visual Masking and Social Interaction Models for Pedestrian Trajectory Prediction

Pedestrian trajectory prediction is critical for autonomous driving, crowd analysis, and urban surveillance. To address insufficient interaction modeling in existing methods, we propose two enhancements. First, a field-of-view(FoV)-aware masking mechanism filters irrelevant interactions by dynamically adjusting to pedestrian distances and motion directions. Second, we introduce a more detailed modeling of the influence among pedestrians, replacing the traditional modeling based on reciprocal of distance. This model, integrated with FoV masks to construct sparse adjacency matrices for graph convolution. A temporal convolutional network then predicts trajectories as bivariate Gaussian distributions. Evaluations on ETH and UCY benchmarks demonstrate the effectiveness of our method.