The paper presents ADM, an accelerated diffusion-based motion prediction model that addresses the limitations of standard diffusion models in terms of high computational cost and sensitivity to noise. The key contributions are:
The paper first encodes the scenario information using a Scenario Encoder module to capture interactions between agents and the environment. Then, the Motion Pattern Estimator learns to model the prior distribution of trajectories by predicting the mean, variance, and navigation nodes. This estimated prior distribution is then refined through a Conditional Diffusion Denoising Module, which iteratively removes noise to generate the final predicted trajectories. The model also includes a Probability Predictor and a Scale Net to estimate the likelihood of each potential action and the scale of the Laplace distribution for the regression loss, respectively.
The proposed ADM framework achieves state-of-the-art performance on the Argoverse motion forecasting dataset, outperforming other methods in terms of minADE, minFDE, and Miss Rate metrics. Additionally, the model demonstrates superior robustness against input noise, maintaining its predictive accuracy even under various levels of disturbance. The key to this performance is the motion pattern estimator, which significantly accelerates the inference time by replacing a large number of denoising steps with a coarse-grained prior distribution estimation, while preserving the representation ability of the diffusion model.
To Another Language
from source content
arxiv.org
Key Insights Distilled From
by Jiahui Li,Ti... at arxiv.org 05-03-2024
https://arxiv.org/pdf/2405.00797.pdfDeeper Inquiries