内容来源：https://research.google/blog/hard-braking-events-as-indicators-of-road-segment-crash-risk/

内容总结：

谷歌研究团队证实：急刹车事件可有效预警路段事故风险

2026年1月13日，谷歌研究院移动人工智能团队负责人Neha Arora与软件工程师Yechen Li共同发布一项研究成果，首次通过大规模实际数据验证了急刹车事件与道路事故风险之间的显著关联。研究指出，利用车载智能系统采集的高频急刹车数据，可实现对交通事故的主动预警，为道路安全管理提供全新思路。

传统交通安全评估主要依赖交警部门的事故统计报告，这类数据虽直接关联人员伤亡与财产损失，但具有明显滞后性，且在主干道和地方道路上事故样本稀疏，往往需要数年时间才能积累足够数据评估特定路段风险。此外，各地事故报告标准不一，进一步增加了建立精准预测模型的难度。

为解决这一问题，研究团队提出将“急刹车事件”作为事故风险的先导指标。急刹车事件指车辆减速度超过特定阈值（-3m/s²）的紧急避险操作，其数据可通过Android Auto等车载平台匿名化采集，覆盖范围广、实时性强。

团队整合了弗吉尼亚州和加利福尼亚州长达十年的公开事故数据与聚合急刹车数据，发现急刹车事件的发生频率与事故率呈显著正相关。在统计模型中，即使控制了交通流量、路段长度、道路类型、坡度、车道变化等复杂因素，急刹车频次较高的路段事故风险依然明显更高。例如，在加州某高速公路汇流段，急刹车发生率是全州平均水平的70倍，该路段在过去十年中平均每六周发生一起事故，急刹车数据成功将其识别为高风险路段，无需依赖长期事故历史记录。

研究表明，急刹车数据的密度远超传统事故报告。在加州和弗吉尼亚州，可观测到急刹车事件的路段数量是报告事故路段数量的18倍，有效填补了道路安全评估中的数据空白。

目前，谷歌研究院移动人工智能团队正与谷歌地图平台合作，将急刹车数据整合至“道路管理洞察”服务中，向交通管理部门提供匿名化、高覆盖、实时性强的路段安全信息，助力其基于先导指标而非滞后事故记录，主动识别并治理高风险路段。

未来，研究团队将进一步优化数据模型，通过对相似路段的聚类分析降低数据稀疏性，推动从风险识别向精准工程干预的转化，例如优化信号灯配时、完善交通标志、 redesign 高风险合流车道几何设计等，系统性提升道路安全水平。

该研究由谷歌与弗吉尼亚理工大学的研究人员合作完成，相关论文《从滞后到先导：验证急刹车事件作为路段事故风险的高密度指标》已公开发表。

中文翻译：

急刹车事件作为路段事故风险的指示指标

2026年1月13日
Neha Arora，移动人工智能团队负责人，与 Yechen Li，软件工程师，谷歌研究院

我们证实了通过 Android Auto 收集的急刹车事件与路段实际事故率之间存在正相关关系。我们确认，急刹车事件发生率较高的道路，其事故风险也显著更高，并建议此类事件可作为道路安全评估的前瞻性指标。

快速链接

交通安全评估传统上依赖于警方报告的事故统计数据，这些数据常被视为"黄金标准"，因为它们直接关联到人员伤亡和财产损失。然而，依赖历史事故数据进行预测建模存在重大挑战，因为这类数据本质上是"滞后"指标。此外，在主干道和地方道路上，事故在统计上属于罕见事件，因此可能需要数年时间才能积累足够数据来为特定路段建立有效的安全状况评估。这种数据的稀疏性，加上各地区报告标准的不一致，使得开发稳健的风险预测模型变得复杂。主动的安全评估需要"前瞻性"指标：即与安全结果相关、但发生频率高于事故本身的风险替代指标。

在《从滞后到前瞻：验证急刹车事件作为路段事故风险的高密度指示指标》一文中，我们评估了急刹车事件作为事故风险可扩展替代指标的有效性。急刹车事件是指车辆向前减速度超过特定阈值（-3m/s²）的情况，我们将其解读为一种紧急避险操作。与基于固定传感器的临近碰撞时间等替代指标不同，急刹车事件源自联网车辆数据，便于进行全网分析。通过将弗吉尼亚州和加利福尼亚州的公开事故数据与 Android Auto 平台提供的匿名、聚合的急刹车事件信息相结合，我们建立了事故（任何严重程度）发生率与急刹车事件频率之间统计学上显著的正相关关系。

数据密度

为了验证该指标的实用性，我们分析了10年的公开事故数据以及聚合的急刹车事件测量数据。急刹车事件的直接优势在于其信号密度。我们对加州和弗吉尼亚州路段的分析显示，观测到急刹车事件的路段数量是报告有事故路段数量的18倍。尽管事故数据众所周知非常稀疏——在某些地方道路上可能需要数年才能观测到一次事件——但急刹车事件提供了连续的数据流，有效地填补了安全地图上的空白。

统计验证

核心目标是确定高频次的急刹车事件是否与高事故率存在因果关系。我们采用了负二项式回归模型（这是《公路安全手册》中的标准方法），以解释比典型事故数据更高的方差程度。

我们的模型结构控制了各种混杂因素，包括：

• 暴露量：交通流量和路段长度。
• 基础设施：道路类型（地方道路、主干道、高速公路）、坡度和累计转弯角度。
• 动态特征：匝道的存在以及车道数量的变化。

结果显示，在两个州，急刹车事件发生率与事故率之间均存在统计学上显著的关联。急刹车频率较高的路段始终表现出较高的事故率，这种关系在不同道路类型（从地方主干道到全封闭高速公路）中都成立。

回归分析还量化了特定基础设施要素的影响。例如，路段上存在匝道与事故风险呈正相关，这很可能是因为并道所需的交织操作所致。

案例研究：高风险并道识别

为了直观展示该指标的实际应用，我们研究了加州连接101号高速公路和880号高速公路的一个高速公路并道路段。历史数据显示，该路段的急刹车事件发生率约为加州高速公路平均水平的70倍，并且在过去十年中平均每六周发生一起事故。

在分析该位置的联网车辆数据时，我们发现其急刹车事件频率在所有路段中排名前1%。急刹车事件信号成功地标记出了这个异常路段，而无需依赖长达十年的事故报告来从统计上确认其风险。这种一致性验证了急刹车事件作为可靠替代指标的效力，即使在没有长期碰撞历史的情况下也能识别高风险位置。

实际应用

将急刹车事件验证为事故风险的可靠替代指标，使原始的传感器指标转变为道路管理中可信赖的安全工具。这一验证支持使用联网车辆数据进行全网交通安全评估，提供更精细的空间和时间粒度。虽然这些结果表明该指标可用于路段风险判定，但并未就与地点无关的驾驶行为风险得出结论。

谷歌研究院的移动人工智能团队正与谷歌地图平台合作，将这些急刹车事件数据集作为"道路管理洞察"服务的一部分对外提供。通过整合这些高密度信号，交通管理机构可以获得聚合的、匿名的数据，与传统事故统计数据相比，这些数据更新鲜，覆盖的道路网络范围也更广。这使得能够使用前瞻性指标来识别高风险位置，而不仅仅是依赖滞后且稀疏的碰撞记录。

未来工作

虽然本研究证实急刹车事件是事故风险的一个稳健的前瞻性指标，但仍有机会进一步完善该信号。我们目前正在研究对同质路段进行空间聚类的机制，以进一步减少数据稀疏性。解决这些局限性将有助于从风险识别转向有针对性的工程改造，利用高密度数据为具体的基础设施干预措施提供信息，范围涵盖信号配时调整、改进标志标线，乃至高风险并道车道的几何线形重新设计。

致谢

此项工作是谷歌和弗吉尼亚理工大学研究人员共同努力的成果。我们感谢合著者 Shantanu Shahane、Shoshana Vasserman、Carolina Osorio、Yi-fan Chen、Ivan Kuznetsov、Kristin White、Justyna Swiatkowska 和 Feng Guo。我们也感谢 Aurora Cheung、Andrew Stober、Reymund Dumlao 和 Nick Kan 在将这项研究转化为实际应用方面所做的贡献。

英文来源：

Hard-braking events as indicators of road segment crash risk
January 13, 2026
Neha Arora, Mobility AI Team Lead, and Yechen Li, Software Engineer, Google Research
We establish a positive association between hard-braking events (HBEs) collected via Android Auto and actual road segment crash rates. We confirm that roads with a higher rate of HBEs have a significantly higher crash risk and suggest that such events could be used as leading measures for road safety assessment.
Quick links
Traffic safety evaluation has traditionally relied on police-reported crash statistics, often considered the "gold standard" because they directly correlate with fatalities, injuries, and property damage. However, relying on historical crash data for predictive modeling presents significant challenges, because such data is inherently a "lagging" indicator. Also, crashes are statistically rare events on arterial and local roads, so it can take years to accumulate sufficient data to establish a valid safety profile for a specific road segment. This sparsity paired with inconsistent reporting standards across regions complicates the development of robust risk prediction models. Proactive safety assessment requires "leading" measures: proxies for crash risk that correlate with safety outcomes but occur more frequently than crashes.
In "From Lagging to Leading: Validating Hard Braking Events as High-Density Indicators of Segment Crash Risk", we evaluate the efficacy of hard-braking events (HBEs) as a scalable surrogate for crash risk. An HBE is an instance where a vehicle’s forward deceleration exceeds a specific threshold (-3m/s²), which we interpret as an evasive maneuver. HBEs facilitate network-wide analysis because they are sourced from connected vehicle data, unlike proximity-based surrogates like time-to-collision that frequently necessitate the use of fixed sensors. We established a statistically significant positive correlation between the rates of crashes (of any severity level) and HBE frequency by combining public crash data from Virginia and California with anonymized, aggregated HBE information from the Android Auto platform.
Data density
To validate the utility of this metric, we analyzed 10 years of public crash data alongside aggregated HBE measurements. The immediate advantage of HBEs is the density of the signal. Our analysis of road segments in California and Virginia revealed that the number of segments with observed HBEs was 18 times greater than those with reported crashes. While crash data is notoriously sparse — requiring years to observe a single event on some local roads — HBEs provide a continuous stream of data, effectively filling the gaps in the safety map.
Statistical validation
The core objective was to determine if a high frequency of HBEs causally links to a high rate of crashes. We employed negative binomial (NB) regression models, a standard approach in the Highway Safety Manual (HSM), to account for a higher degree of variance than is typically found in crash data.
Our model structure controlled for various confounding factors, including:

• Exposure: Traffic volume and segment length.
• Infrastructure: Road type (local, arterial, highway), slope, and cumulative turning angle.
• Dynamics: Presence of ramps and change in the number of lanes.
  The results demonstrated a statistically significant association between HBE rates and crash rates across both states. Road segments with higher frequencies of hard braking consistently exhibited higher crash rates, a relationship that holds true across different road types, from local arterials to controlled-access highways.
  The regression analysis also quantified the impact of specific infrastructure elements. For instance, the presence of a ramp on a road segment was positively associated with crash risk in both states, likely due to the weaving maneuvers required for merging.
  Case study: High-risk merge identification
  To visualize the practical application of this metric, we examined a freeway merge segment in California connecting Highway 101 and Highway 880. Historical data indicates this segment has an HBE rate approximately 70 times higher than the average California freeway, and averaging a crash every six weeks for a decade.
  When analyzing the connected vehicle data for this location, we found that it ranked in the top 1% of all road segments for HBE frequency. The HBE signal successfully flagged this outlier without relying on the decade of crash reports it took to statistically confirm the risk. This alignment validates HBEs as a reliable proxy capable of identifying high-risk locations even in the absence of long-term collision history.
  Real-world application
  Validating HBEs as a reliable proxy for crash risk transforms a raw sensor metric into a trusted safety tool for road management. This validation supports the use of connected vehicle data for network-wide traffic safety assessment, offering enhanced spatial and temporal granularity. While these results indicate utility for road segment risk determination, they do not draw conclusions about location-independent driving behavior risk.
  The Mobility AI team at Google Research is working with Google Maps Platform to externalize these HBE datasets as a part of Roads Management Insights offering. By integrating these high-density signals, transportation agencies can access aggregated, anonymized data that is substantially more fresh and that covers a wider breadth of the road network compared to traditional crash statistics. This allows for the identification of high-risk locations using leading indicators rather than relying solely on lagging and sparse collision records.
  Future work
  While this study confirms that HBEs are a robust leading indicator of crash risk, there are opportunities to further refine this signal. We are currently investigating mechanisms to spatially cluster homogenous road segments to reduce data sparsity even further. Addressing these limitations will enable the transition from risk identification to targeted engineering, where high-density data informs specific infrastructure interventions ranging from signal timing adjustments and improved signage to the geometric redesign of high-risk merge lanes.
  Acknowledgements
  This work was a collaborative effort involving researchers from Google and Virginia Tech. We thank our co-authors Shantanu Shahane, Shoshana Vasserman, Carolina Osorio, Yi-fan Chen, Ivan Kuznetsov, Kristin White, Justyna Swiatkowska, and Feng Guo. We also appreciate the contributions of Aurora Cheung, Andrew Stober, Reymund Dumlao, and Nick Kan in translating this research into practical applications.