Source-linked AI summary
SFMGNet: A Physics-based Neural Network To Predict Pedestrian Trajectories
Sakif Hossain, Fatema T. Johora, Jörg P. Müller, Sven Hartmann, Andreas Reinhardt
TL;DR
Autonomous systems need to predict road users’ motion in mixed traffic, where interactions and limited interpretability remain challenges. SFMGNet combines social-force and group-behavior modeling with an MLP, and the paper reports realistic trajectories and better-than-state-of-the-art performance on ETH and UCY.
Problem
Predicting road users’ trajectories in mixed traffic is challenging because interactions overlap, while limited interpretability remains an obstacle for autonomous systems.
Source: §1 · p. 1Method
Results
SFMGNet predicts realistic trajectories and outperforms state-of-the-art models on the ETH and UCY datasets.
Source: §6 · p. 14Takeaways & Limitations
The authors report that SFMGNet’s predictions can be causally interpreted and may contribute to planning and understanding motion behavior in robots and autonomous vehicles.
Source: §6 · p. 14The SFMG formulation omits a force to avoid further complexity.
Source: §3 · p. 4
Abstract
from arXiv · showhide
Autonomous robots and vehicles are expected to soon become an integral part of our environment. Unsatisfactory issues regarding interaction with existing road users, performance in mixed-traffic areas and lack of interpretable behavior remain key obstacles. To address these, we present a physics-based neural network, based on a hybrid approach combining a social force model extended by group force (SFMG) with Multi-Layer Perceptron (MLP) to predict pedestrian trajectories considering its interaction with static obstacles, other pedestrians and pedestrian groups. We quantitatively and qualitatively evaluate the model with respect to realistic prediction, prediction performance and prediction "interpretability". Initial results suggest, the model even when solely trained on a synthetic dataset, can predict realistic and interpretable trajectories with better than state-of-the-art accuracy.
1. Introduction
Mixed-traffic trajectory prediction must account for overlapping interactions among road users, groups, and static obstacles. This paper proposes a pedestrian-motion model that combines social forces and group behavior with a neural network.
1. Introduction
- In mixed-traffic zones, autonomous systems need trajectories of other road users to plan their own motion and navigate safely.Prediction is complex because road users, groups, and static obstacles interact.Source: §1 · pp. 1–2
- Pedestrian prediction estimates motion over the next few seconds from prior motion and environment-related information.The paper emphasizes pedestrians because they are vulnerable road users and form the majority in most mixed-traffic zones.Source: §1 · p. 2
- The proposed model predicts pedestrian motion while considering static obstacles, other pedestrians, and pedestrian groups.It combines the classical social force model, extended with group behavior, and a neural network.Source: §1 · p. 2
2. Related work
Physics-based and learning-based trajectory models offer complementary strengths, while each has limitations in scalability or interpretability. The paper combines SFMG with a neural network that estimates force-equation parameters in separate modules.
2. Related work
- Physics-based approaches such as the social force model can produce realistic, interpretable motion, but parameter estimation limits scalability and requires expert involvement.These constraints can restrict adaptation to new road-user types.Source: §2 · pp. 2–3
- Neural trajectory models can learn motion dynamics from data, but their sample and model-size demands and black-box behavior limit explainability.The paper notes that explainability matters for analyzing autonomous-system behavior, attributing responsibility, and building trust.Source: §2 · p. 3
- Prior hybrid approaches leave gaps in interpretable force modeling, real-data evaluation, or pedestrian-group interactions.The paper says no cited work considered interactions among pedestrians and pedestrian groups, despite groups contributing toward 70% of road users.Source: §2 · p. 3
- The proposed approach uses SFMG as a prior, with separate neural networks estimating individual forces before combining them into total force.The architecture mimics SFMG equations and is intended to retain social-force interpretability.Source: §2 · p. 3
3. Social force model with pedestrian groups (SFMG)
The SFMG formulation models pedestrian motion through forces that include obstacle and pedestrian interactions, destination attraction, and group behavior. Its group force combines visibility and attraction components.
3. Social force model with pedestrian groups (SFMG)
- Obstacle repulsion represents a pedestrian’s tendency to maintain safe distance, with interaction strength, distance, and direction defined between pedestrian and obstacle.The formulation also describes repulsion between pedestrians and an anisotropic factor for their interaction.Source: §3 · p. 4
- SFMG adds a group force comprising visibility and attraction components to the classical social force model.Visibility represents keeping group members within the pedestrian’s field of view; attraction represents maintaining group coherence.Source: §3 · pp. 4–5
- The group visibility and attraction terms use global strength parameters, desired velocity, and a unit vector between a group member and its centroid.The visibility term concerns keeping other members in view, while the attraction term supports group coherence.Source: §3 · p. 5
- The paper uses these social-force and group-behavior equations to design its pedestrian-motion prediction architecture.Source: §3 · p. 5
4. Physics-based neural network (SFMGNet)
SFMGNet predicts pedestrian motion by estimating the four forces in the SFMG model with separate networks, then recombining them into a total force. Its inputs include observed pedestrian trajectories and an estimated goal direction.
4. Physics-based neural network (SFMGNet)
- Four separate networks estimate acceleration, boundary repulsion, pedestrian repulsion, and group force, while a final network combines their outputs.Figure 2 maps net1–net4 to those forces; net3 contains nine instances. The individual networks are designed to mimic the SFMG force equations.Source: §4 · p. 5, Figure 2 · p. 6, §4.2 · p. 7
4.1. Problem statement
- The problem is to predict future trajectories for all pedestrians from their observed trajectories in the scene.For each pedestrian, the input covers time steps 0 through T, and the predicted future covers T+1 through T_e.Source: §4.1 · p. 5
4.2. Individual modules
- The acceleration module estimates instantaneous and desired velocity from normalized position changes and velocity magnitudes, then combines them to estimate acceleration force.Position inputs are normalized by subtracting the first position in the window to avoid spatial bias.Source: §4.2 · p. 6
- The obstacle module estimates repulsion using the nearest obstacle point, while the pedestrian module uses relative direction, distance, and goal direction to estimate repulsion.The pedestrian module uses up to nine nearby pedestrians and recombines their estimated forces; the group module separately combines visibility and attraction forces.Source: §4.2 · p. 7
4.3. Module: Final recombination
- The final recombination network aggregates outputs from net1, net2, net3, and net4 to predict the pedestrian’s total force.It uses two MLPs, with a ReLU activation in the first; because only the nearest obstacle point is considered, obstacle forces are not summed.Source: §4.3 · p. 8
4.4. Goal prediction
- An Interaction Multiple Model estimator selects the most likely goal-hypothesis trajectory by comparing Kalman-filter trajectories with observations.The estimated goal supplies the goal-directed unit vector used by net1 and net3.Source: §4.4 · p. 8
5. Evaluation
The evaluation examines SFMGNet’s realism, interpretability, and accuracy using synthetic training data and real-world benchmark datasets. It includes feature extraction, a trajectory case study, force analysis, quantitative comparisons, and ablations.
5.1. Experimental setup
- The evaluation uses 1,000 synthetic 30-second simulations, split into training, development, and test sets, and tests on ETH and UCY scenes.The split is 50%, 25%, and 25%, respectively; benchmark trajectories are sampled every 0.4 seconds.Source: §5.1 · pp. 8–9, §5.2 · p. 9
5.3. Feature extraction
- Feature extraction derives obstacle distances and directions, neighboring-pedestrian features, and group-centroid direction and rotation angle from trajectories, group information, and obstacles.The procedure uses the nearest obstacle and up to nine neighboring pedestrians.Source: §5.3 · p. 9, Figure 3 · p. 10
5.4. Realistic trajectory prediction
- In an ETH case study of eight pedestrians, SFMGNet predictions generally resemble real paths and capture subtle deviations, unlike flatter SFMG simulations that may miss destinations.Predictions use 1.2 seconds of observed motion to forecast up to 4.8 seconds.Source: §5.4 · pp. 10–11, Figure 4 · p. 10
5.5. Interpreting trajectory prediction
- For pedestrian 85, predicted total force stays near 0.01 units on each axis, while force components show pedestrian repulsion and group force that SFMG underrepresents.The predicted group force better reflects pedestrian 85’s desire to stay closer to group member 84; the analysis does not decompose group force into its components.Source: §5.5 · pp. 11–12, Figure 6 · p. 12
5.6. Quantitative comparison
- 0.15 ADE and 0.26 FDE are SFMGNet’s average errors, and it outperforms every baseline in every scene.The model is trained solely on synthetic data and predicts the next 4.8 seconds from 1.2 seconds of observed motion.Source: Table 1 · p. 13, §5.6 · p. 13
- SFMGNet predicts 0.0% near-collisions in every reported ETH and UCY scene, using a threshold of 0.1 meter.The authors treat this result as evidence that the predicted trajectories are feasible and realistic.Source: §5.6 · p. 13
5.7. Ablation study
- Adding obstacle repulsion and then pedestrian repulsion improves ablation performance, while the full SFMGNet significantly outperforms the variants without group force.The Att variant uses acceleration force alone; ABr adds obstacle repulsion, and ABrPr also adds pedestrian repulsion.Source: §5.7 · pp. 13–14, Table 2 · p. 14
6. Conclusion and outlook
SFMGNet combines a social force model with group behavior modeling and an MLP; evaluation on ETH and UCY supports realistic trajectories, state-of-the-art performance, and causal interpretability.
6. Conclusion and outlook
- The authors conclude that SFMGNet predicts realistic trajectories and outperforms state-of-the-art models on ETH and UCY.They also report that predictions can be causally interpreted.Source: §6 · p. 14
- SFMGNet combines the social force model with group behavior modeling and an MLP to predict pedestrian trajectories amid static obstacles, other pedestrians, and groups.The authors say the combination lets the MLP inherit the SFMG model’s interpretable nature.Source: §6 · p. 14
- The authors identify adding cars and cyclists, and exploring new road-user types when relevant data is scarce or absent, as directions for further work.Source: §6 · p. 14