AutoZYX
Open Research Plan · DRIVEResearch × AD Safety Joint Lab
用航测自然驾驶数据回答一个很实际的问题：同样是被加塞、跟随前车制动、前车切出暴露障碍物，商用车驾驶员和乘用车驾驶员，是不是在用同一套行为边界？如果不是，差异能不能支撑商用车自动驾驶的参数定义、系统设计和测试用例生成？
A practical question answered with aerial naturalistic driving data: when cut in, when following a braking lead, when a lead vehicle cuts out and reveals an obstacle — do truck drivers and car drivers operate on the same behavioral boundaries? If not, can the difference inform parameter definition, system design and test-case generation for commercial-vehicle autonomous driving?
这项研究不是简单比较“商用车更保守还是乘用车更激进”，而是建立一套按车辆类型、驾驶视野高度、车辆动力学能力和场景簇上下文分层的人类驾驶行为参考模型，用于自动驾驶系统的参数定义、系统设计和测试用例生成。
This is not a simple “trucks are more conservative, cars more aggressive” comparison. It builds a human-driving reference model stratified by vehicle type, driver eye-height, vehicle dynamics capability and scenario-cluster context — to feed parameter definition, system design and test-case generation for automated driving.
加塞 cut-in、前车制动 lead-vehicle deceleration、前车切出 cut-out——恰好是 UNECE R157 附件 3 与 ISO 34502 框架下的典型扰动场景。
Cut-in, lead-vehicle deceleration, and cut-out — exactly the disturbance scenarios in UNECE R157 Annex 3 and the ISO 34502 framework.
乘用车驾驶员与商用车驾驶员。不只比较主车类型，也比较目标车类型、前车类型和前前车可见性。
Passenger-car vs commercial-vehicle drivers — comparing not only ego type, but target-vehicle type, lead-vehicle type and front-front visibility.
输出 P5/P50/P95 分布、收敛状态、场景簇标签、差异显著性、系统设计建议与 OpenSCENARIO 测试用例候选。
P5/P50/P95 distributions, convergence status, scenario-cluster labels, significance of differences, design recommendations and OpenSCENARIO test-case candidates.
现有自动驾驶安全开发里，“人类驾驶员参考行为”常常被当成一个整体使用。但商用车和乘用车差异很大：驾驶员职业属性不同，坐姿和视野高度不同，制动响应链路不同——气压制动有 0.5–1 s 的建立滞后，整车质量大、有效制动距离长——车辆的最大可达减速度能力也不同。
更关键的是，商用车更高的驾驶位可能改变驾驶员能看到的“场景簇”。当前车是乘用车时，商用车驾驶员可以越过前车提前看到前前车、拥堵波和道路形态；当前车本身是大车时，这种视野优势又可能消失，同时近端盲区更大。这种“可见信息集合”的变化，无法只用目标车 TTC 一个指标解释。
In AD safety work, “the human driver reference” is often treated as a single thing. Yet commercial and passenger vehicles differ sharply: driver profession, seating and eye-height, the braking response chain (air brakes add 0.5–1 s of build-up lag, larger mass means longer effective stopping distance), and the maximum achievable deceleration all differ.
More importantly, a truck's higher seat may change the scenario cluster the driver perceives: when the lead is a car, the truck driver can see over it to the front-front vehicle, congestion waves and road geometry earlier; when the lead is itself a large vehicle, that advantage disappears and the near-field blind zone grows. A single target-vehicle TTC cannot capture this.
同一个 ISO 34502 / R157 场景，在不同车辆本体上可能对应不同的人类行为边界。安全阈值可以有统一底线，但拟人化、舒适性和测试代表性，不应只有一套乘用车参数。
The same ISO 34502 / R157 scenario may correspond to different human behavioral boundaries on different vehicle bodies. The safety floor can be unified — but human-likeness, comfort and test representativeness should not rest on a single passenger-car parameter set.
从事故后果和运输暴露量看，商用车不是一个边缘车型变量，而是高后果交通安全问题。2022 年我国道路交通事故造成 60,676 人死亡；2016 年营运车辆肇事统计中，营运货车占事故数的 67.53%、占死亡人数的 74.83%；2024 年一季度我国公路货运量达到 90.1 亿吨。换句话说，研究商用车驾驶行为不是为了做一个“车型对比”的小题目，而是在高暴露量、高后果的运输系统里寻找可用于安全开发的人类行为证据。
From the perspective of crash consequence and transport exposure, commercial vehicles are not a marginal vehicle-type variable but a high-consequence road-safety problem. Road crashes in China caused 60,676 deaths in 2022; in the 2016 operating-vehicle crash statistics, operating trucks accounted for 67.53% of crashes and 74.83% of deaths; and China's road freight volume reached 9.01 billion tons in Q1 2024. In other words, this is not a small “vehicle-type comparison” topic — it seeks human-driving evidence inside a high-exposure, high-consequence transport system.
货车、重型牵引挂车一旦进入事故链条，后果往往比普通乘用车更重。用于商用车自动驾驶的参考行为模型，不能直接照搬乘用车的制动、跟车和变道参数。
Once trucks and tractor-trailers enter an accident chain, consequences are often heavier than passenger-car crashes. A reference model for commercial-vehicle AD cannot simply reuse passenger-car braking, following and lane-change parameters.
17 起较大公路货运事故分析识别出 203 个事故致因、114 个事故要素；事故进程与设备环境之外的上层因素约占 45%，事故参与者中上层参与者占 55%。这提醒我们：驾驶行为要和企业监管、车辆状态、道路环境一起解释。
The 17 major road-freight accident analysis identified 203 causal factors and 114 elements; about 45% of causes sit outside the accident process and equipment/environment layer, and 55% of actors are upper-level actors. Behavior must therefore be interpreted with company supervision, vehicle condition and road context.
货运事故样本和 CRSS 样本都反复出现前车停车 / 制动引发追尾、相邻车辆变道引发擦碰、夜间雨天与坡道等要素。这些信息可以反向校准自然驾驶数据扩样顺序和 OpenSCENARIO 测试优先级。
Both freight-accident and CRSS samples repeatedly surface rear-end crashes after lead stopping/braking, side-swipe events after adjacent-lane changes, and night/rain/grade factors. These can calibrate NDD scaling order and OpenSCENARIO test priorities.
这份规划不是孤立的题目。把团队已有研究资产接上后，它的定位会更准：它是一条主线里的一块拼图，而不是一个一次性成果。
This is not a standalone topic. Connected to the team's existing assets, it is sharper — one module of a larger throughline, not a one-off output.
团队主线是用航测数据 + 标准法规 + 专家知识构建中国驾驶员基础模型（DFM）；商用车研究正是其中“车辆本体差异”的专门模块。
The throughline is a Chinese Driver Foundation Model built from aerial data + standards + expert knowledge; commercial-vehicle behavior is its “vehicle-body difference” module.
传统“本车–单一交互车”关系会丢失复杂交互；商用车视野更高，恰好需要用场景簇而不是单一 TTC 来解释。
The classic “ego–single target” relation loses complex interaction; a truck's higher view is exactly what calls for a scenario cluster rather than a single TTC.
L4 智能重卡的 ODC 定义、SOTIF HARA、触发条件识别、整车级接受准则和仿真测试，正缺“人类参考能力模型”这一层证据，这里补上。
L4 truck programs need ODC definition, SOTIF HARA, triggering-condition identification, vehicle-level acceptance criteria and simulation — and lack the “human reference capability” evidence layer this supplies.
所以更准确的定位不是“商用车与乘用车驾驶行为差异统计”，而是面向商用车自动驾驶安全开发的、车辆类型特化的人类驾驶行为参考模型。统计分布只是第一层，最终要进入 DFM 查询、场景簇生成、SOTIF 触发条件，以及安全 / 舒适 / 效率三类测评。
So the sharper framing is not “statistics of behavioral differences” but a vehicle-type-specific human-driving reference model for commercial-vehicle AD safety development. Distributions are only layer one; they feed DFM queries, scenario-cluster generation, SOTIF triggering conditions, and safety / comfort / efficiency evaluation.
本页采用固定顺序：应对加塞、应对前车制动、应对前车切出。每个场景都同时给出公开研究术语、数据定义标签、过程定义、图示时间点和关键指标，避免“一个场景三套叫法”。
This page uses one fixed order: cut-in response, lead-vehicle deceleration response, and cut-out response. Each scenario shows the public research term, data-definition label, process definition, figure time points and metrics, so the same scenario does not carry three competing names.
Definition: CutInFigure timeline: T0-T5
相邻车道车辆切入本车道。图示时间线把过程点与指标点放在同一条轴上：T2 是最小 TTC，T3 是本车最大减速度。数据表中的 keyframe1–keyframe5 可作为切入过程区间，本研究在这个区间内统计本车峰值制动减速度、TTC 曲线和最小 TTC 发生位置。
An adjacent-lane vehicle merges into the ego lane. The figure places process points and metric points on the same timeline: T2 is minimum TTC, and T3 is maximum ego deceleration. In the data table, keyframe1-keyframe5 define the cut-in interval; within that interval, this study measures peak ego braking, the TTC curve and the min-TTC position.
Definition: DecelerationThreshold: 2 m/s² (0.2g)Metric points: T_min TTC + T_max
同车道前车减速制动。该场景可分为四个阶段：前车减速、后车减速、前车停止减速、后车停止减速，并以 2 m/s²（0.2g）作为减速度阈值。本研究额外关注商用车是否因为更高视野提前获得前前车信息。
The same-lane lead vehicle decelerates. The event is split into four phases: lead decelerates, follower decelerates, lead stops decelerating, and follower stops decelerating, using a 2 m/s² (0.2g) deceleration threshold. This study additionally asks whether commercial vehicles anticipate earlier because of higher eye height.
Definition: CutOutMetrics: TTC1/TTC2 + DHW1/DHW2
前车 TV1 切出，本车前方的主导目标从 TV1 转移到前前车 TV2。图示同时给出 ego-TV1 与 ego-TV2 两组最小 TTC / 最小 DHW；关键指标包括 TTC1（ego-TV1）和 TTC2（ego-TV2）。本研究进一步把“TV2 是否提前可见”作为商用车高视野假设的核心变量。
Lead vehicle TV1 cuts out, and the dominant object ahead of ego switches from TV1 to front-front vehicle TV2. The figure marks minimum TTC and minimum DHW for both ego-TV1 and ego-TV2; key metrics include TTC1 for ego-TV1 and TTC2 for ego-TV2. This study further treats whether TV2 was visible in advance as the key variable for the high-eye-height hypothesis.
视图与关键帧定义来源：自动驾驶安全联合实验室《高速公路典型驾驶场景提取与参数分析》（长春绕城高速航测，飞行高度 200m，60h+ 原始视频，50h 经质检后场景提取）。Diagrams and key-frame definitions: AD Safety Joint Lab, “Highway Typical Driving Scenario Extraction and Parameter Analysis” (Changchun ring-expressway aerial survey, 200 m altitude, 60h+ raw video, 50h after QC).
三个可证伪的假设，把“差异”讲成可检验的机制，而不是模糊的“更保守”。
Three falsifiable hypotheses that frame the “difference” as a testable mechanism, not a vague “more conservative”.
受车辆质量、制动建立时间和舒适性约束，商用车可能倾向更早预判、更低峰值减速度、更长制动持续时间，而不是在更晚时刻用更大减速度处理风险。
Constrained by mass, brake build-up time and comfort, trucks may anticipate earlier and brake with a lower peak over a longer duration — rather than handling risk later with a harder stop.
前车是乘用车时，商用车主车可能提前看到前前车或拥堵波；前车是大车时，这一优势消失。视野收益应按前车类型、车高和道路几何分层。
When the lead is a car, a truck ego may see the front-front vehicle or congestion wave early; when the lead is large, that advantage vanishes. The visibility benefit must be stratified by lead type, vehicle height and road geometry.
只看目标车 TTC，很多“为什么此时制动”解释不了；加入前前车、相邻车、后车压力、车道可用性和遮挡关系后，商用车与乘用车的差异会更清晰。
Target-vehicle TTC alone cannot explain much of “why brake now”; adding the front-front, adjacent and rear vehicles, lane availability and occlusion makes the commercial-vs-passenger difference clearer.
按主车速度分箱、主车类型、切入车类型、交通密度和相对速度分层，比较峰值制动减速度（最小加速度）、制动起点、制动建立时间、jerk、最小 TTC 与最小 TTC 发生时刻。
Stratify by ego speed bin, ego type, cut-in vehicle type, traffic density and relative speed; compare peak braking deceleration (the most negative acceleration), brake onset, build-up time, jerk, min TTC and the timing of min TTC.
同样的 min TTC，对商用车可能对应更早的预判、更低的峰值制动或更大的制动余量；对乘用车可能对应更晚但更强的制动。需要比较 TTC 曲线的形状，而不只取一个最小值。
The same min TTC may, for a truck, mean earlier anticipation, lower peak braking or a larger margin; for a car, later but harder braking. Compare the shape of the TTC curve, not just a single minimum.
比较前车开始制动、前前车开始制动、本车松油门、本车制动、达到最大减速度之间的时间差。如果商用车能提前看到前前车，反应应更早、更平缓。
Measure time gaps between front-front braking, lead-vehicle deceleration, ego throttle release, ego braking and peak deceleration. If trucks see the front-front earlier, their response should be earlier and gentler.
将 cut-out reveal 分成“前前车已可见 / 部分遮挡 / 完全遮挡”三类，分析本车首次响应时刻、障碍物 TTC、制动强度，以及是否存在提前减速。
Split cut-out reveal into front-front “already visible / partly occluded / fully occluded”, and analyze first-response time, obstacle TTC, braking intensity and any anticipatory deceleration.
构建包含目标车、前车、前前车、相邻车、后车、车道、匝道、拥堵波和遮挡关系的场景簇，用多目标上下文解释驾驶员行为——这正是商用车高视野研究的关键增量。
Build a cluster containing target, lead, front-front, adjacent, rear vehicles, lanes, ramps, congestion waves and occlusion relations — explaining behavior with multi-target context, the key increment for high-eye-height vehicles.
将货运事故聚类和关联规则中的高后果场景，与航测自然驾驶中的高暴露场景对齐：前车停车 / 制动追尾对应前车制动，邻道变道擦碰对应加塞，前车切出暴露障碍物对应 cut-out reveal。这样可以先分析“既常见又高后果”的单元，而不是平均用力。
Align high-consequence scenarios from freight-accident clustering and association rules with high-exposure NDD scenarios: lead stop/braking rear-end maps to lead-vehicle deceleration, adjacent-lane side-swipe maps to cut-in, and lead cut-out revealing an object maps to cut-out reveal. This prioritizes units that are both common and consequential.
试点样本边界。下面的数字来自单个 1 小时航测样本（DRIVEResearch_expressway_1h_cutin_scenarios_20260416，长春城市快速路），商用车有效样本只有 50 条（40–60 km/h 段 34 条、60–80 km/h 段 16 条），且速度分布偏低。它们只能作为研究设计的方向性线索，远未达到每个场景 × 速度段 ≥ 500 条的收敛目标，不能当作结论。
Pilot-sample boundary. These numbers come from a single 1-hour aerial sample (DRIVEResearch_expressway_1h_cutin_scenarios_20260416, Changchun urban expressway). The commercial sample is only 50 valid events (34 at 40–60 km/h, 16 at 60–80), skewed to lower speed. They are directional clues for study design only — far below the ≥500-per-cell convergence target, and not conclusions.
| 指标 (40–80 km/h 合并)Metric (40–80 km/h pooled) | 乘用车主车 valid n=530Car ego valid n=530 |
商用车主车 valid n=50Commercial ego valid n=50 |
解读边界How to read it |
|---|---|---|---|
| 主车速度 P50Ego speed P50 | 61.3 km/h | 55.7 km/h | 商用车样本速度偏低，后续必须按速度分箱比较。Commercial sample skews slow — must compare within speed bins. |
| 峰值制动减速度 P50Peak deceleration P50 | −0.568 m/s² | −0.475 m/s² | 商用车中位制动更温和，但样本量与速度都未对齐。Gentler median braking for trucks — but sample size and speed are not matched. |
| 最小 TTC P50Minimum TTC P50 | 16.4 s | 24.7 s | 提示商用车可能更早保持风险余量，也可能只是样本构成差异。Hints at an earlier risk margin for trucks — or just sample composition. |
| 峰值制动发生位置 P50Peak-braking position P50 | 51.0% | 62.2% | 商用车制动峰值更靠后，这一线索值得验证。Truck braking peaks later in the process — worth verifying. |
| 最小 TTC 发生位置 P50Min-TTC position P50 | 60.6% | 62.6% | 两类主车在 min TTC 时间点上接近，需结合 TTC 曲线形状分析。Similar min-TTC timing — must be read together with TTC-curve shape. |
位置 = 关键指标发生帧在加塞过程（keyframe1→keyframe5）中的相对位置。上述数值由原始事件工作簿重算：主车车型取自 ego_veh.class，最小 TTC 与峰值制动的帧取自 feature 表的 TTC_min_frame / Max_deacceration_frame。
Position = where the metric's frame falls within the cut-in process (keyframe1→keyframe5). Values recomputed from the raw workbooks: ego type from ego_veh.class; min-TTC and peak-braking frames from the feature sheet's TTC_min_frame / Max_deacceration_frame.
幅值 (m/s²)，P10 / P50 / P90 · 加塞 40–80 km/hmagnitude (m/s²), P10 / P50 / P90 · cut-in 40–80 km/h
秒 (s)，P10 / P50 / P90 · 加塞 40–80 km/hseconds, P10 / P50 / P90 · cut-in 40–80 km/h
中位位置 (%)，0% = 加塞起点，100% = 完成median position (%), 0% = onset, 100% = complete
曲线形状为示意；标注的最小 TTC 值与位置来自样本curve shape illustrative; marked minima are from the sample
方向性读法：在这个样本里，商用车主车的中位峰值制动更温和（−0.47 vs −0.57 m/s²）、最小 TTC 更大（24.7 vs 16.4 s）、峰值制动更靠后（62% vs 51%）。但这一切都与“速度偏低 + 样本只有 50 条”强烈混杂——所以它是问题，不是答案。
Directional reading: in this sample, commercial egos show gentler median peak braking (−0.47 vs −0.57 m/s²), larger min TTC (24.7 vs 16.4 s), and later peak braking (62% vs 51%). All of it is heavily confounded by lower speed and n=50 — so it is the question, not the answer.
即使是乘用车，驾驶决策也不只基于加塞车、制动前车或切出后的障碍车这一两个目标，而是基于驾驶员视野内所有车辆与道路条件构成的“场景簇”。商用车视野更高、更不易被相邻车辆遮挡，其场景簇包含的车辆更多——因此场景簇更适合用于商用车驾驶行为分析。
Even for a car, the decision is driven not by one or two targets but by the whole scenario cluster of vehicles and road conditions in the driver's field of view. A truck's higher, less-occluded viewpoint admits more vehicles into that cluster — which is why the cluster lens fits commercial-vehicle behavior especially well.
提取方法借鉴本团队“视野引导多车交互场景簇”方法（专利与自然基金在研），核心是四步联动：
Extraction follows our “vision-guided multi-vehicle interaction cluster” method (patent & NSFC in progress), a four-step pipeline:
以本车为原点构建极坐标注意力模型：前方高注意区（±30°，距离随车速自适应 d=v·t_preview）、侧方 UFOV 中注意区（±30–75°）、后视镜低注意区。商用车更高的眼点直接体现为更大的前向注意距离与更小的遮挡——这正是“高视野”进入模型的接口。
A polar attention model around the ego: a high-attention front cone (±30°, speed-adaptive range d=v·t_preview), a side UFOV band (±30–75°), and a low-attention mirror zone. A truck's higher eye-point enters the model directly as a longer forward range and less occlusion.
对过滤后的车辆两两计算纵向 TTC、PET、DRAC 等多类时空安全指标，按阈值构造稀疏交互图，处于同一连通分量的车辆判定为一个场景簇——区别于只用单一 TTC 的二元检测。
Compute longitudinal TTC, PET, DRAC between filtered vehicles, build a sparse interaction graph, and treat each connected component as one cluster — beyond single-TTC binary detection.
按关键交互车进入/退出视野划分场景边界（高速示例：导入 20 m、核心 200 m、导出 80 m），使每个场景簇有物理含义明确、可复现的时序结构。
Segment by when key interacting vehicles enter/leave the field of view (highway example: 20 m lead-in, 200 m core, 80 m lead-out), giving each cluster a reproducible, physically meaningful structure.
把每个场景簇在核心时刻的多车空间分布离散编码，统计哪些多车分布模式在中国交通中真正常见、且风险高——把理论枚举（如 JAMA Annex C）与真实数据统计接上。
Encode the multi-vehicle layout at the core instant and rank which configurations are both common and high-risk in real traffic — connecting theoretical enumeration (e.g. JAMA Annex C) with empirical statistics.
航测“上帝视角”在这里有独特价值：俯视轨迹同时包含所有交互车辆的真值（无遮挡），可以用 S1 的人因视野模型把鸟瞰数据“还原”成以驾驶员为中心的交互图，从而量化“商用车比乘用车多看到了哪些车”。
The aerial “god view” is uniquely valuable here: top-down trajectories hold ground truth for every interacting vehicle (no occlusion), so the S1 human-vision model can re-project bird's-eye data into an ego-centred interaction graph — quantifying exactly which vehicles a truck sees that a car does not.
| 场景Scenario | 图示时间点与数据定义Figure time points and data definition | 行为指标Behavior metrics | 场景簇指标Cluster metrics |
|---|---|---|---|
| 应对加塞Cut-in response | 图示 T0 开始切入；T1 压线；T2 最小 TTC；T3 本车最大减速度；T4 越过车道线；T5 完成。数据表用 keyframe1–5 定义切入区间。Figure: T0 tends to cut in; T1 presses lane line; T2 minimum TTC; T3 maximum ego deceleration; T4 crosses lane line; T5 complete. The data table uses keyframe1-5 to define the cut-in interval. | peak ego braking · brake onset · jerk · min TTC · min-TTC position · TTC slope · THW · gap · cut-in lateral speed · Δv | 相邻车数 · 后车压力 · 目标车前方可用空间 · 目标车类型 · 可换道空间#adjacent · rear pressure · space ahead of target · target type · ego escape room |
| 应对前车制动Lead-vehicle deceleration response | T1 前车开始制动；T2 后车开始制动；T3 前车结束制动；T4 后车结束制动；有效减速度阈值 2 m/s²（0.2g）T1 lead starts braking; T2 ego starts braking; T3 lead ends braking; T4 ego ends braking; effective deceleration threshold 2 m/s² (0.2g) | response delay · brake build-up · peak braking · braking duration · min TTC · min THW · residual gap | 前前车可见性 · 前车类型 · 制动波传播 · 可变道空间 · 曲率/坡度front-front visibility · lead type · brake-wave propagation · escape room · curvature/grade |
| 应对前车切出Cut-out response | T0 开始切出；T1 压线；T2 本车开始转向；T3 TV1 越过车道线；T4 本车开始加速；T5 完成切出；核心指标为 TTC1/TTC2，可扩展到 DHW1/DHW2。T0 tends to cut out; T1 presses lane line; T2 ego starts steering; T3 TV1 crosses lane line; T4 ego starts accelerating; T5 complete; core metrics are TTC1/TTC2, extendable to DHW1/DHW2. | time-to-reveal · TTC1 · TTC2 · min DHW1/2 · response mode · peak braking · steering response · residual clearance | 障碍物类型 · 前车遮挡比例 · 主车眼高代理 · 前前车提前可见时间 · 相邻车道可用性obstacle type · lead occlusion ratio · ego eye-height proxy · front-front lead-time · adjacent-lane availability |
数据审计与车辆类型标注Audit & vehicle-type labeling
关键帧统一映射Map key frames consistently
速度分箱与 ODD 标签过滤Speed bins & ODD filtering
分布、置信区间与收敛检查Distributions, CIs & convergence
车型差异显著性检验Significance testing by type
测试用例与系统设计映射Map to test cases & design
统计结果最终应可被查询：用户问“60–80 km/h 城市快速路、商用车主车被乘用车 cut-in 时，P10 峰值减速度、P10 最小 TTC、峰值制动位置是多少”，系统返回数值、样本量、分箱、收敛状态和适用边界。这让分布从一张静态表，变成 DFM 可调用的能力。
Results should ultimately be queryable: a user asks “at 60–80 km/h on an urban expressway, when a commercial ego is cut in by a car — what are the P10 peak deceleration, P10 min TTC and peak-braking position?”, and the system returns the value, sample size, bin, convergence status and validity boundary — turning a static distribution table into a callable DFM capability.
为不同车辆本体提供人类参考的 TTC、THW、gap、max decel、jerk、response delay 与舒适减速度分布，而不是把乘用车分布直接套到商用车自动驾驶系统。
Vehicle-body-specific human-reference distributions of TTC, THW, gap, max decel, jerk, response delay and comfortable deceleration — instead of porting car distributions onto truck AD.
把商用车高视野与制动能力约束转成感知前向范围、前前车跟踪、遮挡推理、规划提前量与制动建立时间的要求。
Turn high eye-height and braking limits into requirements on forward sensing range, front-front tracking, occlusion reasoning, planning look-ahead and brake build-up time.
用真实分布生成代表性 / 边界 / 挑战三类测试候选；挑战不等于随意极限化，而是从 P5/P10 或事故分布边界抽样。
Generate representative / boundary / challenge cases from real distributions; “challenge” means sampling P5/P10 or crash-distribution edges, not arbitrary extremes.
把“触发条件”从单一目标车参数扩展为场景簇触发条件，例如“前车切出且前前车静止、商用车因视野高度提前可见”的可审查证据。
Extend “triggering conditions” from a single target to cluster-level conditions — e.g. an auditable record of “lead cuts out, front-front stationary, truck sees it earlier due to eye-height”.
下面是给任何想认领这个话题的人的工作分解（周期为指示性）。它本身就是开放的——可以整体认领，也可以只做其中一个工作包或一个速度段。
A work breakdown for anyone picking up the topic (durations indicative). It is itself open — take the whole thing, or just one package or one speed bin.
| WP | 周期Effort | 任务Task | 可交付物Deliverable |
|---|---|---|---|
| WP0 | 2w | 统一车辆类型、时间点、字段字典、脱敏边界与质量规则。Unify vehicle types, time points, field dictionary, anonymization & quality rules. | 指标字典、schema、质量清单metric dictionary, schema, quality checklist |
| WP0.5 | 2w | 把货运事故中的典型场景、关联规则和高风险要素映射到三类自然驾驶场景。Map freight-accident scenarios, association rules and high-risk factors to the three NDD interaction scenarios. | 事故-自然驾驶场景映射表、扩样优先级accident-to-NDD mapping, scaling priorities |
| WP1 | 4w | 完成加塞场景商用车 vs 乘用车扩样分析。Scale up the cut-in commercial-vs-passenger analysis. | 加塞对比证据包、收敛矩阵、首版网页cut-in evidence pack, convergence matrix, first web page |
| WP2 | 6w | 扩展到前车制动与前车切出，补齐关键帧与 TTC 曲线指标。Extend to lead-vehicle deceleration and cut-out; complete key frames and TTC-curve metrics. | 前车制动 / 前车切出证据包、OpenSCENARIO 候选lead-deceleration / cut-out packs, OpenSCENARIO candidates |
| WP3 | 6w | 构建场景簇与可见性代理模型。Build the scenario cluster & visibility proxy model. | 场景簇提取脚本、可见性标签、遮挡分类cluster extraction scripts, visibility labels, occlusion classes |
| WP4 | 4w | 形成商用车自动驾驶开发建议与标准测试映射。Produce truck-AD design recommendations & standard test mappings. | 系统设计建议、测试用例库、落地路线与标准化建议design recommendations, test-case library, deployment path and standardization inputs |
围绕加塞场景，检验商用车与乘用车的人类响应边界是否存在统计显著差异，并转化为车辆本体特化的参考分布。
Test whether commercial and passenger human-response boundaries differ significantly under cut-in, then translate them into vehicle-body-specific reference distributions.
从单目标 TTC 扩展到多车可见性与遮挡关系，解释高驾驶视角下的预判行为，并服务感知范围、规划提前量和测试用例设计。
Extend single-target TTC to multi-vehicle visibility and occlusion, explaining anticipation under high eye-height and informing sensing range, planning horizon and test-case design.
把 P5–P95、收敛检查、场景簇标签与 OpenSCENARIO 连成可复现证据链，最终进入仿真、准入测试和 DFM 数据产品。
Connect P5–P95, convergence checks, cluster labels and OpenSCENARIO into a reproducible chain that feeds simulation, admission testing and DFM data products.
⚠️ 标记处为待核实项（DOI、卷期、R157 数值版本等）；请在正式引用前核对原文。本页数据数字均可由上述加塞试点数据重算复现。
⚠️ marks items to verify (DOIs, volumes, the R157 number version, etc.) before formal citation. Every figure on this page is reproducible from the cut-in pilot data above.
如果你在做商用车自动驾驶、SOTIF、场景测试、自然驾驶数据分析，或者正在负责商用车 AD 产品验证，欢迎基于这个规划继续拆解问题、复现实验、提出新假设。可以整体认领，也可以只做一个工作包、一个场景或一个速度段。数据集可以开放，研究思路也应该开放——真正有价值的不是把题目攥在手里，而是让更多人把问题做深，并把方法落到真实系统和测试流程里。共同验证、工程试点与数据获取可直接联系。
If you work on commercial-vehicle AD, SOTIF, scenario testing, NDD analysis, or product validation for truck AD, take this plan and break it down further, reproduce the experiments, and propose new hypotheses. Claim the whole topic, or just one work package, scenario or speed bin. Datasets can be open; research ideas should be too. The real value is not holding the topic, but letting more people deepen the question and move the method into real systems and test workflows. Reach out for joint validation, engineering pilots and data access.