The study evaluates the performance of various pedestrian trajectory prediction models, including SGAN, Y-Net, Trajectron++, Social-Implicit, and AgentFormer, on the widely-used ETH/UCY dataset. The evaluation focuses on three key aspects:
Accuracy: The models are assessed based on the Average Displacement Error (ADE) and Final Displacement Error (FDE) when generating a single trajectory, which is crucial for practical applications in autonomous driving.
Feature requirements: The impact of the observed motion history on prediction performance is investigated by limiting the number of available timesteps (1, 2, and 8) provided to the models.
Computational efficiency: The inference time of each model is measured to gauge how well they scale with an increasing number of agents in the scene.
The results show that while the constant velocity model (CVM) provides a good approximation of the overall dynamics in many cases, additional features need to be incorporated to reflect common pedestrian behavior. Trajectron++ and Social-Implicit, which leverage graph-based interaction modeling, yield the most accurate results among the investigated architectures. However, the trade-off between accuracy and runtime reveals that the CVM remains the most suitable method for the ETH/UCY dataset, as many learning-based approaches struggle to handle static scenarios and state changes effectively.
The qualitative analysis further highlights the strengths and weaknesses of the different approaches, indicating that future research should focus on developing hybrid methods that can better integrate spatial information, motion history, and intention recognition to address the challenges encountered in real-world autonomous driving scenarios.
Egy másik nyelvre
a forrásanyagból
arxiv.org
Mélyebb kérdések