Human motion trajectory prediction using the Social Force Model for real-time and low computational cost applications

TL;DR

The paper tackles real-time, multimodal human trajectory prediction for social navigation by integrating a Social Force Model–based scene representation with a Generative Adversarial Network and a Conditional Variational Autoencoder (SoFGAN). It introduces a Social Force Representation computed over $M$ angle bins and uses CVAE-driven goal learning to improve long-horizon accuracy, enabling multiple plausible futures. Evaluations on BIWI and UCY (and, in real-world tests, the ETH/UCY/BIWI context) show competitive $mADE$/$mFDE$ scores and notably reduced collision rates, with real-time CPU-based deployment on a ROS-driven Helena robot. The approach offers a low-cost, scalable solution for socially aware motion forecasting suitable for human-robot collaboration and autonomous navigation.

Abstract

Human motion trajectory prediction is a very important functionality for human-robot collaboration, specifically in accompanying, guiding, or approaching tasks, but also in social robotics, self-driving vehicles, or security systems. In this paper, a novel trajectory prediction model, Social Force Generative Adversarial Network (SoFGAN), is proposed. SoFGAN uses a Generative Adversarial Network (GAN) and Social Force Model (SFM) to generate different plausible people trajectories reducing collisions in a scene. Furthermore, a Conditional Variational Autoencoder (CVAE) module is added to emphasize the destination learning. We show that our method is more accurate in making predictions in UCY or BIWI datasets than most of the current state-of-the-art models and also reduces collisions in comparison to other approaches. Through real-life experiments, we demonstrate that the model can be used in real-time without GPU's to perform good quality predictions with a low computational cost.

Figures (5)

• Figure 1: Complex cases. The left picture shows a person who randomly changes the movement direction because is waiting for someone. The right picture shows a bench that can be an obstacle for some pedestrians or a goal for others.
• Figure 2: Social Force Representation. In this example 5 angle bins are considered. The thick vectors are the resultant forces of the thin vectors in each angle bin. There are 2 blue forces of 2 pedestrians and 2 red forces of 2 obstacles applied in a pedestrian located in the center.
• Figure 3: Social Force GAN with CVAE architecture. The GAN generator provides the predictions using the predicted goals given by the CVAE module.
• Figure 4: Predictions when Helena is not moving. The predictions are the red and green lines in the ground. The visualization is obtained using RViz.
• Figure 5: Predictions during an encounter between 4 people and Helena. The predictions are the colored lines in the ground.