TNT: Target-driveN Trajectory Prediction

概要：

Our key insight is that for prediction within a moderate time horizon, the future modes can be effectively captured by a set of target states
【CC】个人理解核心思想是未来的可能性很多，但是可以通过一组预测状态来刻画。 实际就是后面的先预测一些目标点，然后基于这些点来预测轨迹 这么一个思路。这样做的好处是通过目标点来压缩搜索空间。直观来想这个事情，先规划几个目标点然后走过去，还是比较符合直觉的

• It fifirst predicts an agent’s potential target states T steps into the future, by encoding its interactions with the environment and the other agents.
【CC】要想办法将环检信息和对其他车辆交互进行编码，然后去预测目标点
• TNT then generates trajectory state sequences conditioned on targets
【CC】基于目标点生成轨迹
• A fifinal stage estimates trajectory likelihoods and a fifinal compact set of trajectory predictions is selected
【CC】计算生成轨迹的似然值，选一组作为预期
【CC】整体过程跟multipath有一定的区别，multipath是固定撒点然后对点的偏置做预测，而轨迹本身是数据集提取出来的；但整体Intent Uncertainty+Ctrol Uncertainty两阶段估计的思路还在，只是解题的做法不太一样

This is in contrast to previous work which models agent intents as latent variables, and relies on test-time sampling to generate diverse trajectories.
【CC】这里说其他方式是在测试阶段做采样然后生成轨迹，指的是Kmeas对数据集聚类的做法？隐变量的处理方式在后面的有描述

One approach to model the high degree of multimodality is to employ flexible implicit distributions from which samples can be drawn—conditional variational autoencoders (CVAEs) , generative adversarial networks (GANs), and single-step policy roll-out methods .
【CC】对于多隐变量分布进行建模：CVAE，GAN；single-step policy roll-out这个没看过不太清楚是啥，待学习

these models require stochastic sampling from the latent space to obtain implicit distributions at run time. These properties make them less suitable for practical deployment
【CC】这是多隐变量建模的弱点，需要在运行态进行随机采样，这样效率比较差

TNT consists of three stages:
(a) target prediction which proposes a set of plausible targets (stars) among all candidates (diamonds).
【CC】栅格化的进行撒点（菱形点），基于菱形点进行预测（星星）
(b) target-conditioned motion estimation which estimates a trajectory (distribution) towards each selected target,
【CC】基于预测的目标点（星星）进行轨迹的估计
© scoring and selection which ranks trajectory hypotheses and selec