CVPR2021 自动驾驶运动预测挑战赛Argoverse Motion Forecasting Competition

admin管理员组

文章数量:1530053

Challenge Overview

Argoverse™ is a dataset of high-definition maps and sensor data from Argo AI. In 2019, we released this collection publicly to aid the research community in making advancements in key perception and forecasting tasks for self-driving technology. See more details in our CVPR paper.

New! CVPR 2021 Competition

We are pleased to announce a new Motion Forecasting competition to conclude June 13th, 2021. Winning methods will be presented at the CVPR 2021 Workshop on Autonomous Driving. This year's competition is similar to last year's, but with a few tweaks: 

1. Our Motion Forecasting leaderboard will be sorted by “Probabilistic minimum Final Displacement Error概率最小最终位移误差,” which rewards methods that assign higher probabilities to the predicted trajectory closest to the ground truth. First place on the leaderboard will win $2,000.
2. We are interested in techniques that (a) incorporate social context in interesting ways, (b) are computationally lightweight, (c) make creative use of Argoverse's HD maps, or (d) perform especially well on a subset of challenging scenarios that we’ve identified. To reward such methods we will give two $1000 honorable mention prizes. 
3. We require one page or longer pdf reports to be emailed to argoverse-competition@argo.ai by the June 13th deadline so that we can evaluate methods for prizes and CVPR presentations.

Get started

1. To get started, download Argoverse datasets and maps on our website: https://www.argoverse/.
2. Check out the Argoverse API and ask questions (if you have any) on GitHub: https://github/argoai/argoverse-api.
3. The baseline method on the leaderboard, similar to that described in our CVPR 2019 paper, is released at https://github/jagjeet-singh/argoverse-forecasting. We also provide the pre-computed features used in our paper here.

Previous Competitions: Our first Motion Forecasting competition concluded at the NeurIPS 2019 Workshop on Machine Learning for Autonomous Driving in Vancouver, BC, on December 14, 2019. You can see details about the winning approaches here. Our second Motion Forecasting Competition concluded at the CVPR 2020 Workshop on Autonomous Driving. The winning approaches are highlighted in this talk.

轻舟智航冠军方案

据悉，轻舟智航获胜方案基础算法架构将原始高清地图数据组织为Lane graph，使用拓展的GCN建模复杂的车道拓扑关系。

重要贡献

轻舟智航获胜方案最重要的贡献是一种新的远程图注意力机制，用于对高清地图中的车道图进行编码。

这种设计的动机是，因其观察到城市地区的车道可以通过邻居、后继和前任等基本交互类型的组合形成非常复杂的排列，在这些排列中，车道不仅与直接连接的车道有交互，也会与其他不直接连接的车道相互作用。

例如在下图的左子图中，汽车在转弯的车道，但却拐向了不与转弯车道相连的车道中，而在右子图中，车辆行驶中因为当前车道线消失，车辆可能并入左边第一车道，但也存在大量的车辆进入左边第二个或其他车道。

自动驾驶所使用的高精地图虽然分辨率高、细节丰富，但仍然不可能完全覆盖人类司机驾驶行为的各种变化。因此，预测算法要理解人类司机的复杂行为，就需要对高精地图中车道间关系的高层次组合进行语义理解。

传统图神经网络的只考虑直接的拓扑连接关系、通过多层组合扩大感受野的方式是不够的，轻舟智航团队提出的长程图注意力机制有更大的灵活性，可以成功理解复杂的车道互动关系。（目前论文尚在投稿，具体构建方法后期会在论文中详述）。

作者实验发现，使用新的图注意力编码高清地图在所有评价指标上带来了显著的提升。

另外作者的其他重要创新点包括：

1. 将轨迹分类损失函数使用高斯混合模型上的negative log likelihood loss；

2. 通过K-means聚合轨迹模式，构筑了模型集成方法。

以上两项方法虽对ADE和FDE（轨迹预测误差）的影响不大，但显著改进了轨迹概率预测，所以大大提高了最终 minFDE 的 brier score。

本文标签： 挑战赛ArgoversecompetitionForecastingMotion