Socially-Informed Reconstruction for Pedestrian Trajectory Forecasting

TL;DR

This work tackles pedestrian trajectory forecasting under social dynamics by integrating a socially aware forecaster with a social reconstructor to learn robust representations. It introduces pseudo-trajectory augmentation generated from partially reconstructed pasts and a CVAE-based forecaster trained with an ELBO objective, complemented by a social loss to enforce physically and socially viable predictions. The method demonstrates state-of-the-art or competitive performance on ETH/UCY and SDD, yielding more stable predictions as evidenced by KDE analyses. The proposed reconstruction-augmented framework holds practical significance for safer, more reliable motion planning in autonomous systems.

Abstract

Pedestrian trajectory prediction remains a challenge for autonomous systems, particularly due to the intricate dynamics of social interactions. Accurate forecasting requires a comprehensive understanding not only of each pedestrian's previous trajectory but also of their interaction with the surrounding environment, an important part of which are other pedestrians moving dynamically in the scene. To learn effective socially-informed representations, we propose a model that uses a reconstructor alongside a conditional variational autoencoder-based trajectory forecasting module. This module generates pseudo-trajectories, which we use as augmentations throughout the training process. To further guide the model towards social awareness, we propose a novel social loss that aids in forecasting of more stable trajectories. We validate our approach through extensive experiments, demonstrating strong performances in comparison to state of-the-art methods on the ETH/UCY and SDD benchmarks.

Figures (9)

• Figure 1: Overview of our proposed method. During training we continuously generate new socially-aware samples and add them to the training set. At runtime, we only use the encoder and forecaster to predict future timesteps.
• Figure 2: A detailed depiction of our method is presented. Our proposed model uses a social forecaster, a social reconstructor, and a pseudo-trajectory generator to augment the training data. The loss shown here consists of the forecaster loss, the reconstructor loss, and our novel social loss.
• Figure 3: Prediction results for our method compared to two ablated models as well as Agentformer on two examples of the ETH scene. Past and future ground truth trajectories are shown in blue and green dashed lines, while the prediction samples are illustrated with purple dashed lines. 'Ours w/o A' indicates our method without Augmentations. 'Ours w/o A and R' is our method without the augmentations and social reconstructor.
• Figure 4: Sensitivity analysis on the value of $\epsilon$ and its effect on six evaluation metrics: $\text{ADE}^{min}_{20}$, $\text{FDE}^{min}_{20}$, $\text{ADE}^{mean}_{20}$, $\text{FDE}^{mean}_{20}$, KDE, and the overlap count percentage between pedestrians.
• Figure 5: Examples of trajectory prediction with targets in close proximity of each other. Past and future timesteps are denoted by red and blue colors, respectively.