Social Pooling with Edge Convolutions on Local Connectivity Graphs for Human Trajectory Prediction in Crowded Scenes

Human trajectory prediction is a quite challenging task mainly due to numerous social interactions and plausible paths in complex crowed scenarios and varying environments. Data-driven machine learning approaches based on Recurrent Neural Networks (RNNs) have, recently, achieved significant results in modelling human-human interactions in a scene. However, information-sharing pooling modules across RNN Encoders which operate on a local spatial context fail to model long-term scene level correlations, while other that adopt a more global approach are restricted to a rather simplistic formulations due to high computational costs. In this work, we introduce a novel pooling mechanism designed to perform trajectory pooling on a higher semantic level. We have developed a novel multi-layer network architecture based on a new Edge Convolutional operator acting on irregular data which is able to generalize local human-human interactions on a semantic social context. To assess the performance of the proposed social pooling with edge convolutions, we have integrated it into a state-of-the-art trajectory prediction framework based on Generative Adversarial Networks (GANs). Our module managed to overall outperform, by a significant margin, several state-of-the-art pooling modules in real-world challenging benchmark datasets.