轨迹预测Leapfrog Diffusion Model for Stochastic Trajectory Prediction

论文速读

CVPR2023
暂时没有上传的开源代码：https://github.com/mediabrain-sjtu/led

解决什么问题

解决实时性问题，扩散模型用在扩散生成当中，虽然取得了很好的预测效果，但是由于扩散模型预测阶段耗时高，影响了未来轨迹预测的实时性。因此作者采用蛙跳的方式加速这个过程。

解决这个问题的几个关键点

总体架构上面提出了哪些创新

直接使用一个网络跳过扩散模型开始的步骤，并且这个跳过的步骤还比较特殊，跳出的结果直接就是一个轨迹集合了。

这里作者说了三个关键点：
1）这里初始跳跃的时候将直接跳过了噪声到轨迹，跳跃的结果就是一些轨迹，之后再从这些轨迹开始扩散；
2）这种方法可以在预测的过程中让输出的这些K个轨迹彼此之间有感知，知道对方是什么情况。
3）作者设计的跳跃扩散模型和普通的扩散模型，都具有最后的逐步降燥的过程，保证了生成数据的质量。

如何实现蛙跳

直接使用蛙跳进行

如何处理轨迹表达和训练问题

0.Abstract

To model the indeterminacy of human behaviors, stochastic trajectory prediction requires a sophisticated multi-modal distribution of future trajectories. Emerging diffusion models have revealed their tremendous representation capacities in numerous generation tasks, showing potential for stochastic trajectory prediction. However, expensive time consumption prevents diffusion models from real-time prediction, since a large number of denoising steps are required to assure sufficient representation ability. To resolve the dilemma, we present LEapfrog Diffusion model (LED), a novel diffusion- based trajectory prediction model, which provides real-time, precise, and diverse predictions. The core of the proposed LED is to leverage a trainable leapfrog initializer to directly learn an expressive multi-modal distribution of future trajectories, which skips a large number of denoising steps, significantly accelerating inference speed. Moreover, the leapfrog initializer is trained to appropriately allocate correlated samples to provide a diversity of predicted future trajectories, significantly improving prediction performances. Extensive experiments on four real-world datasets, including NBA/NFL/SDD/ETH-UCY, show that LED consistently improves performance and achieves 23.7%/21.9% ADE/FDE improvement on NFL. The proposed LED also speeds up the inference 19.3/30.8/24.3/25.1 times compared to the standard diffusion model on NBA/NFL/SDD/ETH-UCY, satisfying real-time inference needs. Code is available at https: //github.com/MediaBrain-SJTU/LED.

为了模拟人类行为的不定性，随机轨迹预测需要对未来轨迹进行复杂的多模式分布建模。新兴的扩散模型在多种生成任务中展示了巨大的表示能力，显示出对随机轨迹预测的潜力。然而，由于确保足够的表示能力需要大量的去噪步骤，因此扩散模型难以实现实时预测。为了解决这个困境，我们提出了 LEapfrog Diffusion 模型 (LED),这是一种基于扩散的新颖轨迹预测模型，能够提供实时、精确和多样化的预测。LED 的核心在于利用可训练的 leapfrog 初始化器直接学习未来轨迹的表达能力多模式分布，从而跳过大量的去噪步骤，显著加速推断速度。此外，leapfrog 初始化器被训练为适当地分配相关样本，以提供预测的未来轨迹的多样性，显著提高预测性能。在包括 NBA/NFL/SDD/ETH-UCY 四个真实数据集的广泛实验中表明，LED 持续稳定地提高性能，在 NFL 数据集上实现了 23.7%/21.9% 的 ADE/FDE 改进。提出的 LED 还与标准扩散模型相比，在 NBA/NFL/SDD/ETH-UCY 上加速推断 19.3/30.8/24.3/25.1 倍，满足了实时推断需求。代码可在 https://github.com/MediaBrain-SJTU/LED 上获得。

1. Introduction

第一段（介绍轨迹预测这个研究方向）

Trajectory prediction aims to predict the future trajectories for one or multiple interacting agents conditioned on their past movements. This task plays a significant role in numerous applications, such as autonomous driving [5, 24], drones [11], surveillance systems [46], human-robot interaction systems [6], and interactive robotics [21, 26]. Recently, lots of fascinating research progresses have been made from many aspects, including temporal encoding [7, 14, 47, 54], interaction modeling [1, 16, 19, 44, 50], and rasterized pre- diction [12, 13, 27, 49, 55]. In practice, to capture multiple possibilities of future trajectories, a real-world prediction system needs to produce multiple future trajectories. This leads to the emergence of stochastic trajectory prediction, aiming to precisely model the distribution of future trajectories.

轨迹预测旨在根据过去运动的条件预测一个或多个相互作用代理的未来轨迹。这个任务在许多应用中发挥着重要作用，例如自动驾驶 [5,24],无人机 [11],监控系统 [46],人类 - 机器人交互系统 [6],和交互机器人 [21,26]。最近，从多个方面取得了许多有趣的研究成果，包括时间编码 [7,14,47,54],相互作用建模 [1,16,19,44,50],和栅格化预测 [12,13,27,49,55]。在实践中，为了捕捉未来轨迹的多种可能性，一个现实世界的预测系统需要生成多个未来轨迹。这导致了随机轨迹预测的出现，旨在精确地建模未来轨迹的分布。

第二段（前人未来轨迹预测方面的研究有哪些）

Previous works have p