RECO: Rotation Equivariant COnvolutional Neural Network for Human Trajectory Forecasting

Pedestrian trajectory prediction is a crucial task for various applications, especially for autonomous vehicles that need to ensure safe navigation. To perform this task, it is essential to understand the dynamics of pedestrian motion and account for the uncertainty and multi-modality of human behaviors. However, existing methods often produce inconsistent and unrealistic predictions due to their limited ability to handle different orientations of pedestrians. In this paper, we propose a novel approach that leverages the Euclidean group C4 to enhance convolutional neural networks with rotation equivariance. This property enables the networks to learn features that are invariant to rotations, thus maintaining consistent output under various orientations. We present the Rotation Equivariant COnvolutional Neural Network (RECO), a model specifically designed for pedestrian trajectory prediction using rotation equivariant convolutions. We evaluate our model on challenging realworld human trajectory forecasting datasets and show that it achieves competitive performance compared to state-of-the-art methods.