PaperlayerView exact revision on arXiv

Source-linked AI summary

Quantifying Spectral Differences in Vehicle Kinematics Between Production Autonomous and Human-Driven Vehicles Across Driving Scenarios

Peiyi Fang, Xiangyu Li, Yonglin Weng, Ke Ma

arXiv:2609.12609v2cs.RO

TL;DR

Empirical evidence on PAV–HV kinematic differences across real-world scenarios is limited, and time-domain metrics may miss differences in signal dynamics. This study compares spectral features from real-world trajectories across scenarios, finding that differences vary by condition, kinematic variable, and vehicle platform.

  • Problem

    Empirical PAV–HV comparisons across real-world scenarios remain limited, and time-domain summaries may miss differences in signal fluctuation patterns and dynamics.

    Source: §1 · pp. 2–3

  • Method

    The study transforms real-world PAV and HV kinematic signals into spectral features and compares their distributions using kernel density estimation and Wasserstein distance across driving scenarios.

    Source: §1 · p. 4

  • Results

    PAV–HV spectral differences vary by scenario, variable, and platform; speed distinguishes cruising from following, yaw acceleration and jerk distinguish turn direction, and acceleration responds most consistently to weather.

    Source: §5 · p. 25

  • Takeaways & Limitations

    The findings indicate that PAV–HV kinematic differences are not governed by one uniform pattern, supporting evaluation across multiple driving scenarios.

    Source: §5 · p. 25

    The available data cover a limited number of PAV platforms and driving scenarios, with relatively small sample sizes in some conditions.

    Source: §5 · p. 25

Abstract

from arXiv · show

Differences in vehicle kinematic characteristics between production autonomous vehicles (PAVs) and human-driven vehicles (HVs) have been limitedly investigated by empirical studies. Most recent studies rely on simulation-based models, while some further investigate low-level adaptive cruise control (ACC) systems in controlled experiments. These methods commonly adapt some time-domain metrics to characterize PAV-HV differences across limited driving conditions. However, current PAVs equipped with high-level autonomous driving systems generate driving behaviors in a black box using data-driven models. These fundamentally different mechanisms for generating behaviors may produce distinct kinematic characteristics in traffic. More importantly, these time-domain metrics cannot reflect frequency-related traffic dynamics across different driving scenarios. Thus, this study adapted a real-world PAV dataset with four PAV platforms and developed a frequency-domain framework to quantify kinematic differences between PAVs and HVs across diverse driving scenarios, including varying driving states, lighting, weather, and vehicle densities. The framework transforms kinematic signals into the frequency domain and extracts spectral features, and then compares these features between PAVs and HVs based on kernel density estimation and Wasserstein distance. The results reveal clear scenario-dependent PAV-HV spectral differences. Specifically, speed-related differences were consistently smaller during car-following than cruising, while rainy conditions consistently enlarged acceleration-related differences compared with clear conditions. These findings highlight the necessity of multi-scenario evaluations and demonstrate the value of frequency-domain analysis for characterizing PAV-HV kinematic differences under real-world conditions.

1. Introduction

Empirical comparisons of PAVs and HVs remain limited across real-world driving scenarios, and common time-domain metrics do not capture differences in signal fluctuation patterns. This study applies frequency-domain analysis to compare their kinematics across scenarios.

1. Introduction

  • Prior empirical studies have used limited vehicle platforms or scenarios, while simulation and low-level ACC studies do not fully represent current high-level automated driving systems.Public trajectory datasets also face limitations in vehicle-type identification, scenario annotation, and trajectory continuity.Source: §1 · pp. 2–3
  • Time-domain metrics often summarize means, variances, and peaks, but signals with similar statistics can differ in their rates of change and fluctuation patterns.Such metrics therefore do not fully capture dynamic signal characteristics related to traffic oscillations and disturbance propagation.Source: §1 · p. 3
  • Frequency-domain analysis separates slow and rapid signal variations and supports common spectral features for comparing different kinematic signals.It has rarely been applied to quantify PAV–HV kinematic differences.Source: §1 · p. 3
  • The framework compares PAV–HV spectral-feature distributions using real-world trajectories, Fourier transforms, kernel density estimation, and Wasserstein distance across multiple scenario dimensions.The comparisons cover driving states, lighting, weather, and vehicle density.Source: §1 · p. 4
  • PAV–HV spectral differences vary across driving conditions, kinematic variables, and vehicle platforms, motivating comparisons across scenarios rather than a single uniform characterization.Speed shows the clearest cruise–follow distinction; yaw acceleration and jerk distinguish turning direction, while acceleration responds most consistently to weather.Source: §1 · p. 4

2. Dataset and Scenario Extraction

The study reconstructs continuous PAVE trajectories for frequency-domain analysis and organizes comparisons by driving state, lighting, weather, and vehicle density. It uses separate PAV platform groups and pooled HV reference distributions, with some turning conditions represented by relatively few segments.

2.1. Dataset Overview and Trajectory Reconstruction

  • PAVE’s 11-second key-frame segments were insufficient for the longer spectral-analysis trajectories, so continuous records were reconstructed from the underlying GNSS database using vehicle identifiers and timestamps.The dataset includes PAV and HV records, vehicle and driving-mode labels, platform information, scenario annotations, and short trajectory segments.Source: §2.1 · p. 5
  • GNSS records collected at 20 Hz were processed to uniform 10 Hz sampling, and speed and yaw rate were used as primary kinematic inputs.Yaw rate was derived from reconstructed heading angle for turning segments.Source: §2.1 · p. 5

2.2. Scenario Definition and Segment Selection

  • Comparisons were organized by driving state, lighting, weather, and vehicle density, with turning segments retained as complete maneuvers and classified by accumulated heading-change sign.Turning constraints excluded gradual roadway curves and irregular heading fluctuations.Source: §2.2 · p. 6, Table 1 · p. 6
  • Non-turning segments required at least 60 seconds, while some PAV platforms had relatively few left- and right-turn segments.The paper attributes the low turn counts to frequent disengagement of automation during turns; Table 2 reports retained segment counts and cumulative durations.Source: §2.2 · p. 6, Table 2 · p. 7
  • AV trajectories were kept separate for four commercial platforms, while HV trajectories were pooled within each scenario category to form reference distributions.The platform-to-model mapping was anonymized.Source: §2.2 · p. 7

3. Methods

The framework compares PAV and HV spectral-feature distributions across driving states, lighting, weather, vehicle density, platforms, and signals. It constructs kinematic signals, transforms them into frequency-domain features, and assesses sampling uncertainty with bootstrap confidence intervals.

3. Methods

  • Processed trajectory segments are grouped by scenario, and their spectral-feature distributions are compared across PAV platforms, scenarios, signals, and features.The framework uses KDE-overlap and Wasserstein-based scores for these comparisons.Source: §3 · p. 7, Figure 2 · p. 8

3.1. Kinematic Signal Construction

  • Speed is smoothed before differentiation to construct speed, acceleration, and jerk signals; turning segments instead use yaw rate, yaw acceleration, and yaw jerk.A centered three-sample moving average reduces small speed-measurement fluctuations.Source: §3.1 · p. 8

3.2. Frequency-Domain Transformation

  • Speed is detrended before spectral transformation, while acceleration and jerk remain undetrended; the signals are mean-centered, Hann-windowed, and Fourier-transformed.The retained frequency range uses a minimum-cycle criterion and extends to the 5 Hz Nyquist frequency.Source: §3.2 · pp. 9, 11

3.3. Spectral Feature Extraction

  • Seven features characterize spectral frequency location, dispersion, cumulative-power distribution, and allocation across low, middle, and high frequency bands.The cumulative-frequency features mark where 20% and 80% of cumulative spectral power are reached.Source: §3.3 · pp. 11–12

3.4. PAV–HV Difference Measures

  • KDE overlap and normalized Wasserstein scores compare unpaired PAV and HV feature distributions, with Wasserstein separation scaled by HV variability.Feature combinations with zero or undefined HV standard deviation are excluded; higher normalized Wasserstein scores indicate greater similarity.Source: §3.4 · pp. 12–14

3.5. Robustness Assessment

  • Bootstrap confidence intervals assess sampling uncertainty by resampling feature samples within each scenario–signal–vehicle combination.The procedure uses 1,000 resamples and percentile-based 95% intervals.Source: §3.5 · p. 14

4. Results

Results compare PAV–HV spectral differences by platform across driving states, turns, lighting, weather, and vehicle density. The reported patterns vary by signal, feature, and scenario, with several environmental analyses focused on AV-4.

4.1. Overview of Scenario-Dependent PAV–HV Kinematic Differences

  • Driving-state results cover all four PAV platforms, while the main-text lighting, weather, and density analyses focus on AV-4.Complete platform-specific results across scenario dimensions, signals, and features are reported in the appendix.Source: §4.1 · p. 14

4.2. Cruise and Car-Following Conditions

  • Car-following generally yields higher SKDE and SW scores for speed and acceleration than cruising, with the contrast strongest for speed and limited for jerk.The pattern is clearest for AV-1 and AV-4; AV-2 and AV-3 show weaker or less consistent changes.Source: §4.2 · p. 16, Figure 4 · p. 16, Figure 5 · p. 17
  • Cruise–follow feature patterns vary across platforms: car-following generally improves agreement for several speed and acceleration features on AV-1 and AV-4, but less consistently on AV-2 and AV-3.The feature-level comparisons are shown for SKDE and SW across platforms.Source: §4.2 · p. 16, Figure 4 · p. 16, Figure 5 · p. 17
  • SKDE and SW trends broadly agree but differ in magnitude; AV-4 bootstrap resampling supports stable speed and acceleration increases under car-following, with jerk relatively unchanged.The bootstrap analysis reports 95% confidence intervals from trajectory-segment resampling.Source: §4.2 · pp. 16–18, Table 5 · p. 18

4.3. Left-turn and Right-turn Maneuvers

  • Turn direction affects higher-order yaw more than yaw rate: right turns generally score higher for yaw acceleration and, less consistently, yaw jerk.The yaw-acceleration difference is particularly evident for AV-3 and AV-4; bootstrap results show comparable yaw-rate scores and higher right-turn scores for yaw acceleration and jerk.Source: §4.3 · pp. 19–20, Table 6 · p. 21

4.4. Lighting and Weather Conditions

  • Lighting effects are feature-dependent: SKDE stays relatively close across day and night-lit conditions, while SW varies more, especially for acceleration and jerk.For speed, the relative ranking of lighting conditions varies across frequency-domain features.Source: §4.4 · p. 21, Figure 10 · p. 22
  • Rain generally lowers SKDE and SW scores than clear weather, while speed discrepancies persist at mid- and high-frequency features and SW changes more for speed and acceleration.The signal-level comparisons are reported for AV-4.Source: §4.4 · pp. 21, 23, Figure 11 · p. 22

4.5. Vehicle-Density Conditions

  • Increasing density is associated with lower acceleration scores and larger differences; speed effects concentrate in low- and mid-frequency features, while jerk stays relatively stable.Wide bootstrap confidence intervals for several density comparisons, especially SW, mean these variations should not be treated as a uniform density effect.Source: §4.5 · pp. 23–24

5. Conclusion

The framework quantifies PAV–HV spectral differences across scenarios using real-world data, Fourier-transformed kinematic signals, KDE, and Wasserstein distance. Results vary by scenario, kinematic variable, and vehicle platform, with stable quantification supported across measures and bootstrap intervals.

5. Conclusion

  • The framework applies Fourier transforms, KDE, and Wasserstein distance to compare PAV–HV spectral features across driving states, lighting, weather, and vehicle density.It is implemented and evaluated using the PAVE dataset, which contains trajectories and driving-scenario annotations.Source: §5 · p. 24
  • KDE and Wasserstein measures show broadly consistent trends, while bootstrap confidence intervals support the statistical reliability of the results.These findings support stable quantification across major driving scenarios rather than dependence on a single distributional measure.Source: §5 · pp. 24–25
  • PAV–HV spectral differences vary in magnitude and direction across driving conditions, kinematic variables, and vehicle platforms.Across platforms, speed most consistently distinguished cruising from following, yaw acceleration and yaw jerk distinguished left from right turns, and acceleration responded most consistently to weather.Source: §5 · p. 25
  • The study covers a limited number of PAV platforms and driving scenarios, with relatively small sample sizes in some conditions.It quantifies kinematic differences but does not explain how they relate to ADS decision-making strategies.Source: §5 · p. 25

CRediT authorship contribution statement

The contribution statement assigns methodology, software, data, analysis, supervision, and writing responsibilities across the listed authors.

CRediT authorship contribution statement

Appendix A. Segment-Duration Statistics

Appendix A reports trajectory-segment duration statistics across driving-state, lighting, weather, and vehicle-density conditions.

Appendix A. Segment-Duration Statistics

Appendix B. Complete Scenario–Signal–Feature Scores

Appendix B provides complete SKDE and SW scores for cruise, car-following, turning, lighting, weather, and vehicle-density conditions across PAV platforms, kinematic signals, and frequency-domain features.

Appendix B. Complete Scenario–Signal–Feature Scores

Appendix C. Complete KDE Distributions

Appendix C presents complete KDE distributions for seven spectral features across driving states, turning conditions, lighting, weather, and vehicle densities.

Appendix C. Complete KDE Distributions

Loading 2609.12609v2…