The Simpler the Better: Constant Velocity for Pedestrian Motion Prediction

Christoph Schöller, Vincent Aravantinos, Florian Lay, Alois C. Knoll

Abstract

Pedestrian motion prediction is a fundamental task for autonomous robots and vehicles to operate safely. In recent years many complex approaches based on neural networks have been proposed to address this problem. In this work we show that - surprisingly - a simple Constant Velocity Model can outperform even state-of-the-art neural models. This indicates that either neural networks are not able to make use of the additional information they are provided with, or that this information is not as relevant as commonly believed. Therefore, we analyze how neural networks process their input and how it impacts their predictions. Our analysis reveals pitfalls in training neural networks for pedestrian motion prediction and clarifies false assumptions about the problem itself. In particular, neural networks implicitly learn environmental priors that negatively impact their generalization capability, the motion history of pedestrians is irrelevant and interactions are too complex to predict. Our work shows how neural networks for pedestrian motion prediction can be thoroughly evaluated and our results indicate which research directions for neural motion prediction are promising in future.

Introduction

Predicting where a pedestrian will walk in the next few seconds is a critical challenge for autonomous vehicles and social robots. To address this, researchers have developed increasingly complex deep learning architectures, including LSTMs, GANs, and attention mechanisms designed to model social interactions and complex motion histories. However, these models often struggle to generalize to new environments. This paper investigates whether the added complexity of these neural networks is truly beneficial or if a simple physics-based baseline can achieve superior performance.

Methodology

The authors evaluate several models against a baseline Constant Velocity Model (CVM). The CVM calculates a velocity vector from the last two observed positions and projects it linearly. A stochastic variant, OUR-S, adds random angular noise to the predictions to allow for a fair comparison with generative models like Social-GAN (S-GAN).

The neural competitors include standard Feed-Forward (FF) networks, LSTMs, and state-of-the-art models such as RED, SR-LSTM, S-GAN, and SoPhie. Evaluation was conducted on the ETH and UCY datasets using a "leave-one-out" cross-validation strategy to test generalization across different scenes.

The Environmental Prior Trap

A major finding is that neural networks often learn "environmental priors"—scene-specific walking patterns—rather than general motion dynamics. For instance, if a training set predominantly features horizontal movement, the network will struggle in a new environment where vertical movement is common.

Comparison of trajectory distributions in the 'Hotel' scene versus other datasets, illustrating the environmental prior trap.

As shown in the figure above, the "Hotel" scene has distinct vertical patterns compared to the "Others" group. Neural models trained on the latter fail on the former because they attempt to apply the horizontal priors they memorized, leading to significant errors until forced to use relative coordinates or data augmentation.

Redundancy of History and Social Interactions

The study also challenges common assumptions about input data:

  1. History is Redundant: Analysis showed that the most recent timestep accounts for over 80% of the predictive influence. Long motion histories are so highly correlated with the immediate past that they provide little additional signal.
  2. Social Interactions are Too Complex: Even when provided with ground-truth future positions of neighbors, neural models failed to improve. This suggests that current architectures cannot effectively reason about the vast state-space of human social navigation.

Comparative Results

The simple CVM baseline, particularly the noisy OUR-S version, consistently outperformed complex neural models. Quantitatively, OUR-S achieved an Average Displacement Error (ADE) of 0.28m, surpassing S-GAN (0.32m) and SoPhie (0.54m).

Visual comparison of predictions from OUR-S, S-GAN, and RED.

As illustrated in the comparison above, the OUR-S model produces a spread of predictions that effectively covers the ground truth, often more accurately than the more complex S-GAN or RED models. Simple Linear Regression also outperformed LSTMs, as its higher bias makes it less prone to overfitting to scene-specific priors.

Conclusion

The success of the Constant Velocity Model serves as a critique of current deep learning approaches in pedestrian motion prediction. It highlights that many complex models are essentially memorizing maps rather than learning motion. Future research should focus on incorporating explicit semantic maps, developing more diverse datasets, and finding more robust ways to model social interactions beyond simple state pooling.

Summary updated

Abstract

Pedestrian motion prediction is a fundamental task for autonomous robots and vehicles to operate safely. In recent years many complex approaches based on neural networks have been proposed to address this problem. In this work we show that - surprisingly - a simple Constant Velocity Model can outperform even state-of-the-art neural models. This indicates that either neural networks are not able to make use of the additional information they are provided with, or that this information is not as relevant as commonly believed. Therefore, we analyze how neural networks process their input and how it impacts their predictions. Our analysis reveals pitfalls in training neural networks for pedestrian motion prediction and clarifies false assumptions about the problem itself. In particular, neural networks implicitly learn environmental priors that negatively impact their generalization capability, the motion history of pedestrians is irrelevant and interactions are too complex to predict. Our work shows how neural networks for pedestrian motion prediction can be thoroughly evaluated and our results indicate which research directions for neural motion prediction are promising in future.
BibTeX
@article{scholler2020simpler,
  author = {Christoph Schöller and Vincent Aravantinos and Florian Lay and Alois C. Knoll},
  title = {{The Simpler the Better: Constant Velocity for Pedestrian Motion Prediction}},
  journal = {IEEE Robotics Autom. Lett.},
  year = {2020},
  doi = {10.1109/lra.2020.2969925},
}
Figure