Lane Departure Assessment via Enhanced Single Lane-Marking

Yiwei Luo, Ping Li, Gang Shi, Zuowei Liang, Lei Chen, Fengwei An2022 · DOI

TL;DR

This paper introduces a novel Lane Departure Warning System (LDWS) that simplifies calibration and improves accuracy by relying on just one lane marking. By employing a precise 3D camera imaging model and a new criterion for assessing yaw angle and distance, the system achieves high accuracy in lane departure detection, even in challenging scenarios.

What Problem Does This Paper Address?

Modern vehicles increasingly incorporate Advanced Driver Assistance Systems (ADAS) to enhance safety. A crucial component of ADAS is the Lane Departure Warning System (LDWS), designed to alert drivers when their vehicle drifts out of its lane, a common cause of accidents due to inattention or inexperience. Existing LDWS often require complex calibration procedures and rely on detecting both lane markings, which can be problematic in adverse weather or when markings are obscured. This paper addresses the need for a more robust, simpler-to-calibrate, and accurate LDWS that can function effectively even when only one lane marking is visible.

Methods

The core of this work lies in an enhanced LDWS that utilizes a novel camera calibration strategy and a refined lane departure assessment criterion.

  • Simplified Camera Calibration: The system employs a 3D camera imaging model that requires only three parallel and equally spaced lines on the ground for calibration. This significantly simplifies the setup compared to methods needing specific geometric arrangements or calibration objects like chessboards. The calibration process deduces the three rotation angles for transforming the camera's coordinate system to the world's coordinate system. Crucially, the camera height and lane width are calculated during this process, avoiding manual measurement errors.
  • Enhanced Lane Departure Assessment: The proposed LDWS uses a criterion that estimates both the vehicle's yaw angle (deviation from the road direction) and the distance to the lane marking. This is achieved by detecting only one lane marking in the image. This "single lane-marking" approach makes the system more resilient to situations where one marking is obscured. The system leverages the calibrated extrinsic camera parameters (rotation angles and height) to derive these critical parameters.

Datasets

The paper evaluates its proposed method using:

  • Highway and Urban Road Image Sequences: These were captured by a camera mounted on a car, with the car's position and orientation varied to simulate real-world driving conditions.
  • KITTI Odometry Dataset: This dataset is a standard benchmark for visual odometry and SLAM, providing ground truth for camera poses. The paper notes that while KITTI provides camera deviation angles, it lacks real-world distance measurements to lane markings, making it suitable for evaluating the yaw angle component but not the distance parameter.

Key Insights and Ideas

  • The Power of "One": The most significant insight is that reliable lane departure warning can be achieved using information from just one lane marking, drastically simplifying the system's robustness and applicability.
  • Calibration Simplicity: The proposed calibration method, requiring only three parallel lines, is a major practical advantage, making it easier to implement in various vehicle setups without stringent installation requirements.
  • Calculated vs. Measured Parameters: Deriving camera height and lane width through calibration, rather than manual measurement, eliminates a significant source of error and improves overall system accuracy.
  • Unified Approach for Straight and Curved Roads: While the paper primarily focuses on straight lanes, it proposes using the tangent line of a curve as an equivalent to a "straight lane-marking" for curved roads, offering a unified approach to lane departure warning.

Key Results

The experimental results demonstrate the effectiveness of the proposed LDWS:

  • High Accuracy: The system achieved an average accuracy of 98.95% on lane departure assessment in experiments.
  • Low Error Rates:
    • Average error for yaw angle ($\theta_y$) was approximately 1 degree.
  • Average error for distance to lane marking ($x_x$) was less than 5 cm.
  • Excellent Warning Rates: The correct warning rate was consistently high across different road types, reaching 98.17% for highway, 98.64% for urban roads, and 100% for the KITTI dataset (using only $\theta_y$ as the criterion).
  • Comparison with State-of-the-Art: The proposed method outperformed several existing algorithms in terms of correct warning rate, particularly those requiring two lane markings. It also showed comparable performance to another 3D imaging model-based approach ([25]), highlighting the advantage of combining image information with a road model.
  • Camera Height and Lane Width Accuracy: Experimental results for camera height and lane width showed average errors of 1.50% and 2.27%, respectively.

Limitations and Future Work

  • Curved Roads: The paper acknowledges that for curved roads, the accuracy of the lane departure warning is influenced by the camera's field of view (FOV). A wider FOV is needed to minimize estimation errors, especially when the road curvature is significant. The difficulty in precisely detecting the tangent line and the difference between the tangent point and the actual closest point on the lane marking (x-point) contribute to errors.
  • Future Work: The authors suggest future research could focus on more advanced algorithms to estimate the x-point more precisely on curved roads. They also propose detecting and utilizing the curved lane-markings directly, rather than just their tangent lines, for a more complex but potentially more accurate approach to curved lane departure warning.

Summary updated

BibTeX
@misc{luo2022lane,
  author = {Yiwei Luo and Ping Li and Gang Shi and Zuowei Liang and Lei Chen and Fengwei An},
  title = {{Lane Departure Assessment via Enhanced Single Lane-Marking}},
  year = {2022},
  doi = {10.3390/s22052024},
}
Figure