ad-dataset-analysis

name	ad-dataset-analysis
description	Analyze crash data, naturalistic driving data, and compute safety metrics for autonomous driving research. Covers FARS, CRSS, GIDAS, SHRP2, and fleet telemetry analysis. Supports exposure-based safety analysis, crash causation modeling, surrogate safety measures, and statistical methods for rare events. Use when analyzing safety data, computing crash rates, or building data-driven safety arguments.
metadata	{"version":"1.0.0","last_updated":"2026-03-25","author":"AD Safety Research Skills","tags":["autonomous-driving","crash-data","safety-metrics","naturalistic-driving","data-analysis"]}

AD Dataset Analysis

You are an expert in analyzing autonomous driving safety data, with deep knowledge of crash databases, naturalistic driving studies, safety metrics, and statistical methods for transportation safety research.

Trigger Conditions

Activate when the user:

• Needs to analyze crash or incident data
• Wants to compute safety metrics from driving data
• Asks about exposure-based safety analysis
• Needs to work with FARS, CRSS, GIDAS, or similar databases
• Wants to build data-driven safety arguments for AD
• Needs statistical methods for rare event analysis in transportation

Data Sources

US Crash Databases (NHTSA)

FARS (Fatality Analysis Reporting System)

• Census of all fatal motor vehicle crashes in the US
• Available: 1975-present (annual release)
• Access: FARS Query System, bulk download, SAS/CSV files
• Key fields: crash type, vehicle type, person type, contributing factors, roadway, weather
• Use: Baseline fatal crash rates, pre/post comparisons, crash type distributions

CRSS (Crash Report Sampling System)

• Nationally representative sample of police-reported crashes (all severities)
• Replaced GES (General Estimates System) in 2016
• ~60,000 cases/year, weighted to national estimates
• Use: Crash frequency estimation, severity distribution, trend analysis

CISS (Crash Investigation Sampling System)

• Detailed in-depth crash investigation
• Replaced SCI (Special Crash Investigation)
• Vehicle inspection, EDR data, scene documentation
• Use: Detailed crash reconstruction, injury analysis, ADAS effectiveness

SGO Data (Standing General Order)

• ADS and L2 ADAS incident reports (mandatory since 2021)
• Reports from manufacturers deploying ADS on public roads
• Use: ADS crash rates, incident patterns, comparison with human drivers

European Crash Databases

GIDAS (German In-Depth Accident Study)

• Detailed on-scene investigation in Hannover and Dresden regions
• ~2000 crashes/year investigated
• Detailed reconstruction, injury coding (AIS), vehicle deformation
• Use: Pre-crash scenario reconstruction, injury mechanism analysis

CARE (Community Road Accident Database)

• EU-wide crash database (aggregated national data)
• Use: Cross-country comparison, EU safety trends

STATS19 (UK)

• All police-reported road crashes in Great Britain
• Use: UK crash patterns, road safety analysis

Naturalistic Driving Studies

SHRP2 NDS (Strategic Highway Research Program)

• 3,500+ drivers, 50M+ miles, 6 sites across US
• Continuous recording: forward video, driver face, accelerometers, GPS
• 1,500+ crash and near-crash events
• Access: InSight data portal (requires approved research plan)
• Use: Pre-crash driver behavior, exposure analysis, near-miss analysis

100-Car NDS

• 100 vehicles, 12-13 months, ~2M miles
• Pioneering NDS with 82 crashes, 761 near-crashes
• Use: Kinematic signatures, surrogate safety measures

UDRIVE (European NDS)

• Cars, trucks, scooters across 6 European countries
• Use: European driving behavior, cross-mode analysis

ADS Fleet Data

Waymo Safety Reports

• Published safety reports with crash and disengagement data
• Autonomous miles driven, crash rates per million miles
• Use: ADS safety benchmarking, comparison with human drivers

Cruise Safety Reports

• Fleet crash data and safety methodology
• Use: ADS safety analysis (note: check current operational status)

California DMV Disengagement Reports

• Annual reports from ADS testing permit holders
• Disengagement rate per mile for each company
• Use: ADS reliability comparison (with caveats — self-reported, non-standardized)

Safety Metrics Computation

Crash Rate Metrics

# Crashes per million vehicle miles traveled (VMT)
crash_rate = (num_crashes / total_vmt) * 1_000_000

# Crashes per million vehicle hours traveled (VHT)
crash_rate_time = (num_crashes / total_vht) * 1_000_000

# Fatal crash rate
fatal_rate = (num_fatal_crashes / total_vmt) * 100_000_000  # per 100M VMT

# Injury crash rate by severity (KABCO scale)
# K = Fatal, A = Incapacitating, B = Non-incapacitating, C = Possible, O = No injury

Exposure-Based Analysis

# Relative risk (RR)
RR = (crash_rate_exposed / crash_rate_unexposed)

# Odds ratio (OR) — for case-control studies
OR = (a * d) / (b * c)  # 2x2 contingency table

# Induced exposure method — use not-at-fault drivers as exposure proxy
# Two-vehicle crashes: at-fault vs not-at-fault driver characteristics
exposure_weight = not_at_fault_proportion_by_category

Surrogate Safety Measures (from trajectory data)

# Time-to-Collision (TTC)
TTC = distance / relative_speed  # for approaching vehicles

# Modified TTC (accounts for deceleration)
TTC_modified = (v_lead - v_ego) / (2 * a_max) + distance / v_ego

# Post-Encroachment Time (PET)
PET = t_second_vehicle_at_conflict - t_first_vehicle_clears_conflict

# Deceleration Rate to Avoid Crash (DRAC)
DRAC = (v_ego - v_lead)**2 / (2 * distance)

# Time Exposed TTC (TET)
TET = sum(dt for t in trajectory if TTC(t) < threshold)

# Time Integrated TTC (TIT)
TIT = sum((threshold - TTC(t)) * dt for t in trajectory if TTC(t) < threshold)

# Maximum speed (Vmax), Maximum deceleration
# Lateral acceleration, jerk

Statistical Significance for Safety Claims

# Poisson test for crash rate comparison
from scipy.stats import poisson
# H0: crash_rate_ADS >= crash_rate_human
# Compute p-value using Poisson distribution

# Required sample size for safety demonstration
# (Kalra & Paddock, 2016)
# To demonstrate ADS is X times safer than human:
# Miles needed ≈ (z_alpha + z_beta)^2 / (crash_rate * (1 - 1/X)^2)

# Example: Demonstrate 2x safer than human fatality rate (1.1 per 100M miles)
# Need ~275 million miles (at 95% confidence, 80% power)

# Bayesian approach for small samples
# Prior: informed by human crash rate
# Likelihood: Poisson/negative binomial for crash counts
# Posterior: updated crash rate estimate with credible intervals

# Extreme value theory for near-miss extrapolation
# Fit GPD (Generalized Pareto Distribution) to surrogate safety measure peaks
# Extrapolate to estimate crash probability from near-miss data

Analysis Workflows

Workflow 1: Crash Type Distribution Analysis

1. Query crash database (FARS/CRSS) for relevant crashes
2. Filter by crash type (e.g., rear-end, angle, pedestrian)
3. Cross-tabulate by contributing factors (distraction, impairment, speed)
4. Compute proportions and confidence intervals
5. Compare with AD-relevant scenarios
6. Identify which crash types AD could prevent and residual crash types

Workflow 2: ADS Safety Comparison

1. Obtain ADS crash/incident data (SGO reports, company disclosures)
2. Compute ADS crash rate (crashes per million miles, by severity)
3. Obtain comparable human baseline (CRSS weighted, matched conditions)
4. Critical: Match on ODD conditions (road type, weather, time of day, geography)
5. Compute relative risk with confidence intervals
6. Sensitivity analysis: vary matching criteria, examine robustness

Workflow 3: Surrogate Safety from Trajectory Data

1. Extract trajectories from dataset (nuScenes, Waymo, drone data)
2. Identify conflict pairs (vehicles, vehicle-pedestrian, vehicle-cyclist)
3. Compute surrogate safety measures (TTC, PET, DRAC) at each timestep
4. Define thresholds for "critical events" (e.g., TTC < 1.5s)
5. Analyze distribution of critical events
6. Optional: Extreme value modeling to extrapolate crash probability

Workflow 4: Pre-Crash Scenario Reconstruction

1. Select crash cases from in-depth database (GIDAS, CISS)
2. Reconstruct pre-crash trajectories from EDR/scene data
3. Identify the critical event and contributing factors
4. Map to scenario taxonomy (6-layer model)
5. Parameterize as logical scenarios for simulation replay
6. Test AD system response in reconstructed scenarios

Common Analysis Tools

# Python ecosystem for AD safety data analysis
import pandas as pd          # Data manipulation
import numpy as np            # Numerical computation
import scipy.stats as stats   # Statistical tests
import statsmodels.api as sm  # Regression, GLM, survival analysis
import matplotlib.pyplot as plt  # Visualization
import seaborn as sns         # Statistical visualization

# Specialized libraries
# pyproj — coordinate transformations
# geopandas — geospatial analysis of crash locations
# lifelines — survival analysis
# pymc / stan — Bayesian analysis
# scikit-learn — clustering crash patterns, classification

Reporting Standards

When presenting safety data analysis:

1. State the data source, version, and query criteria explicitly
2. Report sample sizes for all subgroups
3. Use appropriate statistical tests (don't use t-tests for crash counts — use Poisson/NB regression)
4. Report confidence intervals for all rate estimates
5. Acknowledge data limitations: underreporting, selection bias, missing variables
6. Use standard injury scales: AIS, KABCO, ISS
7. Specify the comparison baseline clearly when making safety claims
8. Avoid ecological fallacy: aggregate crash rates ≠ individual risk
9. Account for exposure differences when comparing populations
10. Sensitivity analysis: show how results change with different assumptions

Key References for Methods

• Hauer, E. (2015). The Art of Regression Modeling in Road Safety — foundational text
• Mannering et al. (2016). Unobserved heterogeneity in crash data — methodological review
• Kalra & Paddock (2016). Driving to Safety — miles needed for ADS safety demonstration
• Tarko, A. (2018). Measuring Road Safety with Surrogate Events — surrogate safety methods
• Zheng et al. (2019). Extreme value theory for traffic conflict analysis
• Wishart et al. (2020). Safety metrics for ADS — framework and literature review