1.预测
1.1光栅化
代表性论文
- Motion Prediction of Traffic Actors for Autonomous Driving using Deep Convolutional Networks (Uber)
- MultiPath (Waymo)
问题
- 渲染信息丢失
- 感受野有限
- 高计算复杂度
1.2图神经网络
1.2.1 图卷积
- LaneGCN (uber 2020)
1.2.2 边卷积
- VectorNet (waymo 2020)
注意:Vectornet的子图使用的是边缘卷积,大图使用的是自注意力机制
Transformer
- mmTransformer (2021)
- WayFormer
- QCNet (2023 CVPR,还没看懂)
1.3基于锚点
- DenseTNT
1.4生成式模型
- Social gan: Socially acceptable trajectories with generative adversarial networks (CVPR 2018)
- Tranjectron++
1.5Metrics
- b-minFDE:(1-p_i)^2 * minFDE_i
- b-minADE:(1-p_i)^2 * minADE_i
参考:https://eval.ai/web/challenges/challenge-page/454/evaluation
2.端到端
2.1光栅化
- End-to-end Interpretable Neural Motion Planner (uber 2019)
- P3 (uber)
2.2图神经网络
2.3Transformer
- UniAD
- VAD
- SparseAD / SparseDrive
2.4采样
-
Curve-based sampler (NMP )
-
Lane-based sampler (P3 uber 2020)
-
Retrieval-based sampler (MP3)
-
如何理解agent centric