Driver's Preview Modeling Based on Visual Characteristics through Actual Vehicle Tests

Hongyu Hu, Ming Cheng, Fei Gao, Yuhuan Sheng, Rencheng Zheng2020 · DOI

TL;DR

This paper introduces a novel method for modeling driver preview behavior by analyzing actual vehicle test data. By combining eye-tracking with vehicle kinematics, the researchers developed a preview model that accurately captures how drivers anticipate the road ahead, offering a more humanized approach to intelligent vehicle control.

What Problem Does This Paper Address?

Intelligent vehicles aim to mimic human driving behavior for smoother and safer control. A critical aspect of human driving is "preview," where drivers use visual information to anticipate upcoming road conditions and adjust their actions accordingly. Existing driver models often rely on simulators, which can lack the fidelity of real-world driving, or use simplified visual models. This paper addresses the need for a more accurate and robust driver preview model derived from actual driving data, specifically focusing on how visual characteristics like fixation points influence this preview behavior across various driving conditions.

Methods

The core of this research lies in a multi-step process to extract and model driver preview behavior:

  • Actual Vehicle Testing: Eight drivers participated in real-world driving tests on straight roads and curves with varying radii (200m, 800m, 1500m) at different speeds (50, 70, 90 km/h).
  • Data Acquisition: Eye movement data (using a head-mounted eye tracker) and vehicle kinematics/dynamics were collected simultaneously. A forward-looking camera captured road scenes.
  • Fixation Point Extraction: The Identification-Deviation Threshold (I-DT) algorithm was employed to distinguish fixation points (where the driver focuses for a period) from scan points (rapid eye movements).
  • Image Processing Pipeline:
    • Distortion Correction: Images from the eye tracker were corrected for lens distortion to ensure accurate spatial representation.
  • Image Matching: The SURF (Speeded Up Robust Features) algorithm was used to match images from the eye tracker with those from a fixed-view driving recorder. This compensates for head movements and ensures a stable reference frame.
  • Preview Point Determination: Fixation points were projected into a world coordinate system. The "preview point" was identified as the center of the distribution of these projected fixation points.
  • Statistical Analysis and Modeling:
    • The Jarque-Bera test was used to confirm that fixation point distributions followed a normal distribution.
  • Parameters of the normal distribution (mean and standard deviation) were fitted for different driving conditions.
  • These parameters were then used to build preview models, specifically probability density maps of preview time, as a function of speed and road curvature.

Datasets

The study utilized data collected from actual vehicle tests involving eight drivers. The dataset comprises:

  • Eye movement data (viewpoints, fixation points, scan points).
  • Forward-looking road scene images.
  • Vehicle kinematics and dynamics data (steering wheel angle, throttle, brake pressure, yaw rate, accelerations).

This dataset was collected under controlled experimental conditions (specific speeds, road curvatures) and serves as the basis for the developed preview models. No pre-existing public datasets were used for training or evaluation; the data was generated specifically for this research.

Key Insights and Ideas

  • Real-world data is crucial: The study highlights the limitations of simulators and emphasizes the value of actual vehicle tests for developing realistic driver models.
  • Fixation points reveal preview behavior: The distribution and characteristics of driver fixation points directly correlate with their preview behavior and anticipation of the road ahead.
  • Normal distribution of fixation: Fixation points, when analyzed longitudinally, tend to follow a normal distribution, allowing for statistical modeling.
  • Speed and curvature are key drivers: Both vehicle speed and road curvature significantly influence the driver's preview time and the dispersion of their gaze. Higher speeds generally lead to longer preview times, while tighter curves shorten them.
  • A unified preview model: The research successfully integrates these factors to create a general preview model applicable to both straight and curved roads within a defined range of speeds and curvatures.

Key Results

  • The developed I-DT algorithm effectively separates fixation points from scan points with high accuracy.
  • A robust image processing pipeline (distortion correction and SURF matching) was established to obtain accurate fixation point coordinates in a fixed perspective.
  • The Jarque-Bera test confirmed the normal distribution of fixation points across various driving conditions.
  • Preview models (probability density maps of preview time) were generated for straight lanes and curves, showing how preview time varies with speed and curvature.
  • Model validation showed high correlation coefficients (r² = 0.965 for straight lanes, r² = 0.932 for curves) between model fitting and analytical calculations, indicating satisfactory performance.

Limitations and Future Work

  • Limited driving scenarios: The current models are applicable to specific speed ranges (50-90 km/h) and curvature radii (200-1800m). More complex scenarios like motorways, slopes, varying road adhesion, and maneuvers (lane changing, overtaking) require further investigation.
  • Generalizability: While the models are designed to be general, their applicability to higher speeds or tighter curves may be limited and requires specific experimental analysis.
  • Physiological characteristics: The authors suggest future work could involve integrating physiological characteristics of drivers to further enhance the humanized design of the preview model.
  • Broader maneuvers: Extending the study to include more complex driving maneuvers would improve the model's adaptability.

Summary updated

BibTeX
@misc{hu2020driver,
  author = {Hongyu Hu and Ming Cheng and Fei Gao and Yuhuan Sheng and Rencheng Zheng},
  title = {{Driver's Preview Modeling Based on Visual Characteristics through Actual Vehicle Tests}},
  year = {2020},
  doi = {10.3390/s20216237},
}
Figure