跳转至

PhaseShift：信号交叉口之间的拓扑感知数据同构化与模型整合

文章背景与核心概要

传统的交通行为模型通常针对单个交叉口进行独立训练，这导致了孤立的模型组合，无法在不同地点之间共享或利用先验证据。为了突破这一局限，作者团队推出了 PhaseShift——一个专为拓扑感知设计的框架，旨在将异构的路侧轨迹统一转换为共享的以交通参与者为中心的表示形式。通过训练单一可复用的骨干网络，PhaseShift 在去除特定场地传统约束的同时，保留了对行为至关重要的拓扑结构。

该研究在佛罗里达州两个地区的五个交叉口进行了评估，采用回放条件下的最佳采样轨迹（best-of-sampled-trajectory）协议。评估结果表明，单一的集成模型（pooled model）在性能上显著优于本地训练方案，不仅降低了长视野预测误差，还展现出强大的零样本（zero-shot）和少样本（low-data）自适应能力。

📌 Summary / 摘要与总结

Traffic-behavior models are traditionally trained independently for individual intersections, creating isolated portfolios of models that cannot share or leverage evidence across different sites.

Traffic-behavior models are traditionally trained independently for individual intersections, creating isolated portfolios of models that cannot share or leverage evidence across different sites.

传统上，交通行为模型都是针对各个交叉口独立训练的，这形成了孤立的模型组合，无法在不同地点之间共享或利用证据。

To overcome this limitation, the authors introduce PhaseShift, a topology-aware framework designed to harmonize heterogeneous roadside trajectories into a shared actor-centric representation. By training a single reusable backbone, PhaseShift removes site-specific conventions while retaining behaviorally relevant topology.

To overcome this limitation, the authors introduce PhaseShift, a topology-aware framework designed to harmonize heterogeneous roadside trajectories into a shared actor-centric representation. By training a single reusable backbone, PhaseShift removes site-specific conventions while retaining behaviorally relevant topology.

为了克服这一局限性，作者引入了 PhaseShift，这是一个拓扑感知框架，旨在将异构的路侧轨迹同构化（harmonize）为共享的、以参与者为中心的表示。通过训练单一的可复用骨干网络，PhaseShift 消除了特定地点的常规限制，同时保留了与行为相关的拓扑结构。

The evaluation—conducted across five intersections in two Florida regions using a replay-conditioned, best-of-sampled-trajectory protocol—demonstrates that a single pooled model significantly improves performance over local training regimes, reducing long-horizon errors and providing robust zero-shot and low-data adaptation capabilities.

The evaluation—conducted across five intersections in two Florida regions using a replay-conditioned, best-of-sampled-trajectory protocol—demonstrates that a single pooled model significantly improves performance over local training regimes, reducing long-horizon errors and providing robust zero-shot and low-data adaptation capabilities.

该评估在佛罗里达州两个地区的五个交叉口进行，采用了回放条件下的最佳采样轨迹协议。结果表明，单一的集成模型显著提升了相较于本地训练机制的性能，降低了长视野误差，并具备了稳健的零样本和少样本自适应能力。

🧭 Abstract / 摘要

Learned traffic-behavior models are commonly trained separately for each intersection, creating model portfolios that cannot share evidence across sites. We present PhaseShift, a topology-aware framework that harmonizes heterogeneous roadside trajectories into a shared actor-centric representation and trains one reusable backbone.

Learned traffic-behavior models are commonly trained separately for each intersection, creating model portfolios that cannot share evidence across sites. We present PhaseShift, a topology-aware framework that harmonizes heterogeneous roadside trajectories into a shared actor-centric representation and trains one reusable backbone.

学习型交通行为模型通常针对每个交叉口分别训练，这导致各个地点之间无法共享证据的模型组合。我们提出了 PhaseShift，这是一个拓扑感知框架，它将异构的路侧轨迹同构化为共享的以参与者为中心的表示，并训练一个可复用的骨干网络。

Ego-relative coordinates, trajectory-induced movement paths, normalized signal context, and variable-cardinality interaction tokens remove site conventions while preserving behaviorally relevant topology. The backbone supports pooled operation, zero-shot at a held-out intersection, and low-data adaptation.

Ego-relative coordinates, trajectory-induced movement paths, normalized signal context, and variable-cardinality interaction tokens remove site conventions while preserving behaviorally relevant topology. The backbone supports pooled operation, zero-shot at a held-out intersection, and low-data adaptation.

自车相对坐标、轨迹诱导的运动路径、归一化的信号上下文以及可变基数（variable-cardinality）交互标记消除了特定地点的惯例，同时保留了行为相关的拓扑结构。该骨干网络支持集成操作、在未见交叉口（held-out intersection）上的零样本泛化以及少样本自适应。

We evaluate five intersections in two Florida regions on balanced field data, 100k training windows and equal-sized test sets per site under a replay-conditioned, best-of-sampled-trajectory protocol. At 10s, one pooled model lowers both minADE and minFDE relative to trained local models at all five sites, with median reductions of 36.8% and 22.0%. Leave-one-intersection-out deployment, including one cross-region fold, beats local training on both 10-s metrics at four of five sites, although short-horizon performance is less uniform.

We evaluate five intersections in two Florida regions on balanced field data, 100k training windows and equal-sized test sets per site under a replay-conditioned, best-of-sampled-trajectory protocol. At 10s, one pooled model lowers both minADE and minFDE relative to trained local models at all five sites, with median reductions of 36.8% and 22.0%. Leave-one-intersection-out deployment, including one cross-region fold, beats local training on both 10-s metrics at four of five sites, although short-horizon performance is less uniform.

我们在平衡的现场数据上对佛罗里达州两个地区的五个交叉口进行了评估，每个站点包含 10 万个训练窗口和等大小的测试集，采用回放条件下的最佳采样轨迹协议。在 10 秒时，与所有五个站点的本地训练模型相比，单个集成模型降低了 minADE 和 minFDE，中位数降幅分别为 36.8% 和 22.0%。留一交叉口部署（Leave-one-intersection-out deployment，包括一个跨区域折叠）在五个站点中的四个站点的两个 10 秒指标上均优于本地训练，尽管短视野性能的均匀性稍逊。

Fine-tuning with 1,000 target update windows improves on zero-shot at three sites and is the strongest regime at one. At site 7, every cross-site mixture sharply lowers long-horizon error under a fixed 100k-window budget; test-likelihood gains argue against a best-of-sample dispersion-only explanation. Local models fall behind calibrated IDM at the two highest-flow sites after long autoregressive rollouts; pretrained-backbone regimes do not.

Fine-tuning with 1,000 target update windows improves on zero-shot at three sites and is the strongest regime at one. At site 7, every cross-site mixture sharply lowers long-horizon error under a fixed 100k-window budget; test-likelihood gains argue against a best-of-sample dispersion-only explanation. Local models fall behind calibrated IDM at the two highest-flow sites after long autoregressive rollouts; pretrained-backbone regimes do not.

使用 1,000 个目标更新窗口进行微调在三个站点上改进了零样本性能，并在其中一个站点成为了表现最强的方案。在站点 7，在固定的 10 万窗口预算下，任何跨站点混合都能大幅降低长视野误差；测试似然增益排除了仅用最佳采样离散度（dispersion-only）来解释的可能。在长时间自回归展开后，本地模型在流量最大的两个站点上落后于校准后的智能驾驶员模型（IDM），而预训练骨干网络方案则没有出现这种情况。

Within this five-site evaluation, PhaseShift demonstrates consolidation across heterogeneous physical control settings while identifying sites that still require adaptation. The protocol measures conditional single-vehicle generation under replayed context, not closed-loop traffic simulation.

Within this five-site evaluation, PhaseShift demonstrates consolidation across heterogeneous physical control settings while identifying sites that still require adaptation. The protocol measures conditional single-vehicle generation under replayed context, not closed-loop traffic simulation.

在这项包含五个站点的评估中，PhaseShift 展示了对异构物理控制设置的整合能力，同时识别出了仍需自适应的站点。该协议测量的是回放上下文下的条件单车生成（conditional single-vehicle generation），而非闭环交通仿真。