Danger period setting 危险阶段设定

[Yiru-Jiao]

Why is it set that

This danger period is considered to end in 0.5 seconds after impact time or at the annotated end time if it is earlier.

in Section 5.2 Ground truth of safety-critical events? Could the time moment when the subject driver reacted be utilised to set a danger period?

This question is asked by Xudong Ren from Southeast University via email.

为什么在第5.2章Ground truth of safety-critical events中，使用这样的设定

This danger period is considered to end in 0.5 seconds after impact time or at the annotated end time if it is earlier.
这里是否可以使用 driver first reacts 作为danger period？

本问题由来自东南大学的任旭东通过邮件提出。

[Yiru-Jiao]

• Setting of danger period end
  The original data in SHRP2 NDS provides an annotation of start time, reaction time, impact time, and end time for each safety-critical event. The end time annotation could be seconds long after the impact time, because the annotation remarks the end of event, but not necessarily the end of a danger period before collision. Therefore, if the end time is within 0.5 seconds after impact time, we use the originally annotated end time as danger period end; otherwise, we use 0.5 seconds after impact time.

• Use of subject driver reaction time
  We don't suggest using the annotated reaction time to define the danger period in an event. If we see each driver having their individual risk perception and reaction model, the reaction time annotations across all events are sampled from thousands of models inconsistently. If we see all drivers collectively as a human risk perception and reaction model, the annotations are resulted from a human model, which is not objective and does not ensure more accurateness than an artificial model.

• 危险期结束的设定
  SHRP2 NDS的原始数据为每个安全关键事件提供了_开始时间_、反应时间、_影响时间_和_结束时间_的标注。结束时间标注可能在影响时间之后长达几秒，因为该标注是事件结束的时间点，而不一定是碰撞前危险期结束的时间。因此，如果结束时间在碰撞时间之后 0.5 秒以内，我们使用原始标注的结束时间作为危险期结束；否则，我们使用碰撞时间之后 0.5 秒作为危险期结束。

• 驾驶员自身反应时间的使用
  我们不建议使用标注的反应时间来定义事件中的危险期。如果考虑每个驾驶员都有各自的风险感知和反应模型，那么跨所有事件的反应时间标注是从成千个模型中不一致地采样得到的。而如果我们将所有驾驶员整体视为一个人类风险感知和反应模型，那么这些标注就源自于一个人类模型。这个模型并非是客观的，也不能保证比人工模型更准确。

[xudongren-seu]

Thank you very much for your detailed reply.

I have another question regarding the definition of the danger period. Would it be appropriate to use the timing of the lead vehicle’s braking or steering maneuver as the start of the danger period? My concern is that using a fixed-length time window before the crash to define the danger period may not fully account for the dynamic evolution of risk across different events. Specifically, the onset of danger can vary depending on how quickly a hazardous situation develops, which could be reflected by the lead vehicle’s critical maneuver (such as braking or sudden lane change).

Would incorporating such event-specific operational cues (e.g., lead vehicle braking or steering) into the definition of the danger period provide a more accurate reflection of risk evolution, compared to simply using a uniform pre-crash time window? I would appreciate your thoughts on this approach.

Thank you again for your insights.

[Yiru-Jiao]

We set the start of the danger period in each event based on the annotated start time in the original SHRP2 NDS. Those annotations were made by human assistants in the SHRP2 project, and indeed vary instead of being fixed.

For more details on how the start time annotations were made, we refer to the report that describes the creation of SHRP2 NDS database: J. M. Hankey, M. A. Perez, and J. A. McClafferty. Description of the SHRP 2 Naturalistic Database and the Crash, Near-Crash, and Baseline Data Sets. Tech. rep. Virginia Tech Transportation Institute, 2016.

For your convenience, we copy the definition of Event Start in the report:

The point in the video when the sequence of events defining the occurrence of the incident, near-crash, or crash begins. Defined as the point at which the Precipitating Event begins (see Precipitating Event [V8]). Value is a timestamp, in milliseconds after the start of the file. NOTE: for cases in which the origin of the Precipitating Event is not visible in the video (e.g., "Other vehicle ahead - stopped on roadway more than 2 seconds" or "Pedestrian in roadway"), the start point for the Precipitating Event would be when the event is first visible in the forward view of the subject vehicle. ALSO NOTE: For baseline events, the Event Start is defined as 1 second (1,000 timestamps) prior to the end of the baseline event.

[xudongren-seu]

Thank you again for your insights.

[xudongren-seu]

I misunderstand at least 4.5s as 4.5s, sorry for that.

[Yiru-Jiao]

Thanks for paying attention to this paper 🙏