跳转至

PhaseShift：信号交叉口之间的拓扑感知数据对齐与模型整合

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，这导致了各个路口模型各自为政，无法在不同站点之间共享或利用经验证据。为了突破这一局限，本文作者推出了 PhaseShift——一个专为拓扑感知的框架，旨在将异构的路侧轨迹统一转换为共享的以参与者为中心的表征。通过训练单一可复用的骨干网络，PhaseShift 能够消除特定地点的布局差异，同时保留与行为相关的拓扑特征。

该研究在佛罗里达州两个地区的五个交叉口进行了评估，采用了基于重放条件（replay-conditioned）和最佳采样轨迹（best-of-sampled-trajectory）的评测协议。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，不仅降低了长程预测误差，还展现出强大的零样本（zero-shot）和少样本（low-data）自适应能力。

📌 摘要

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

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.

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.

交通行为模型通常针对每个交叉口分别训练，从而形成了无法在不同站点间共享证据的孤立模型组合。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协调为共享的以参与者为中心的表征，并训练一个可复用的骨干网络。

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.

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的交互令牌（tokens），消除了地点的特定惯例，同时保留了与行为相关的拓扑结构。该骨干网络支持集成操作、留出交叉口的零样本泛化以及少样本自适应。

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.

我们在平衡的现场数据上评估了佛罗里达州两个地区的五个交叉口，每个站点拥有 10 万个训练窗口和等大小的测试集，采用重放条件下的最佳采样轨迹协议。在 10 秒时，单一集成模型在所有五个站点上均降低了 minADE 和 minFDE，相比于本地训练模型，中位降幅分别达到 36.8% 和 22.0%。留一交叉口部署（包括一个跨地区折叠）在五个站点中的四个站点上击败了本地训练的指标，尽管短期性能表现不太统一。

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.

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结果，并且在一个站点中成为表现最强的方案。在站点 7 中，在固定的 10 万窗口预算下，任何跨站点混合配置都大幅降低了长程预测误差；测试似然度的提升反驳了仅用最佳样本离散度来解释的观点。在长自回归展开后，本地模型在流量最大的两个站点上落后于经过校准的 IDM（智能驾驶员模型），而预训练骨干网络方案则没有。

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.

在此次五站点评估中，PhaseShift 展示了跨异构物理控制设置的模型整合，同时识别出了仍需自适应的站点。该协议测量的是重放上下文下的条件单车生成，而非闭环交通仿真。

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.

🔗 链接与资源