TL;DR
This paper presents a robust vehicle localization system for autonomous vehicles by fusing GNSS, IMU, and DMI sensors with a novel fault-detection mechanism to handle GNSS jumps. It further enhances lateral accuracy by incorporating LiDAR-based curb detection and fitting, significantly improving localization precision in complex urban environments.
What Problem Does This Paper Address?
Precise and reliable localization is a cornerstone for autonomous driving, enabling vehicles to navigate safely and accurately. Traditional approaches often rely on GNSS (Global Navigation Satellite System), but these systems are susceptible to signal blockages and multipath interference, leading to "jumps" in position estimates that can be meters off. Inertial Measurement Units (IMUs) provide continuous data but drift over time, and Distance Measuring Instruments (DMIs) also accumulate errors. While fusing these sensors (GNSS/IMU/DMI) improves robustness, GNSS jumps remain a significant challenge. Furthermore, even with fused sensor data, lateral localization accuracy can be insufficient for safe navigation, especially in urban settings with curbs and lane markings. This paper aims to address both the robustness against GNSS jumps and the accuracy of lateral positioning.
Methods
The core of the proposed approach is a multi-constraint fault-detection mechanism integrated into an Unscented Kalman Filter (UKF)-based sensor fusion framework.
- UKF-based GNSS/IMU/DMI Fusion: The paper utilizes an Unscented Kalman Filter (UKF) for sensor fusion. UKF is chosen over Extended Kalman Filter (EKF) for its ability to handle nonlinearities without requiring Jacobian matrix computations, thus preserving second-order accuracy. The state vector includes position, velocity, orientation (as a unit quaternion), and biases for accelerometers and gyroscopes.
- Multi-constraint Fault Detection for GNSS: To mitigate GNSS jumps, the authors implement a fault-detection system. This system employs several checks:
- Zero-Velocity Update: If the vehicle is stationary (detected by encoders), GNSS measurements are not used to prevent updates to a fixed location.
- Number of Satellites: A minimum number of visible satellites (four or more) is required.
- Dilution of Precision (DOP): HDOP and VDOP values exceeding a threshold lead to discarding measurements.
- Statistical Test (Chi-squared): The innovation (difference between predicted and measured state) and its covariance are used to statistically test for abrupt jumps.
- Position Change Assessment: The change in horizontal position is compared against vehicle speed from encoders.
-
Altitude Component Check: Large deviations in altitude measurements are flagged. If any of these checks fail, the GNSS measurement is rejected, preventing it from corrupting the state estimate.
-
LiDAR-based Lateral Correction: To improve lateral accuracy, the system uses a 3D LiDAR to detect curbs.
- Curb Detection: A multi-step algorithm identifies curb-like points from LiDAR point clouds, filtering out noise and separating obstacles based on height.
- Curb Line Fitting: The RANSAC (Random Sample Consensus) algorithm is employed to fit either straight lines or curves to the detected curb points, robustly handling outliers.
- Lateral Error Estimation: The lateral distance from the vehicle to the detected curb is calculated. This is compared with the lateral distance to the curb defined in the Route Network Definition File (RNDF). The difference represents the lateral localization error.
- 1D Kalman Filter for Lateral Adjustment: A one-dimensional Kalman filter is used to smooth and estimate this lateral error. The output of this filter is then used to adjust the vehicle's global position, correcting the lateral offset.
Datasets
The paper evaluates its approach using a publicly available dataset proposed in [30], specifically the "Campus-0L" dataset. This dataset provides ground truth poses with centimeter-level accuracy. The authors note that the dataset does not include encoder data, so artificial encoder data was generated from RTK GNSS speed measurements to serve as ground truth.
Key Insights and Ideas
- GNSS Jumps are a Major Hurdle: The paper clearly demonstrates that even with sophisticated sensor fusion, GNSS jumps can significantly degrade localization accuracy, leading to erroneous trajectories.
- Multi-Constraint Fault Detection is Crucial: A single check is insufficient; combining multiple statistical and geometric constraints provides a more robust way to identify and reject faulty GNSS measurements.
- LiDAR for Lateral Precision: LiDAR's ability to accurately detect physical road features like curbs offers a powerful way to correct lateral localization errors that are difficult to resolve with GNSS/IMU/DMI alone.
- RANSAC for Robust Curb Fitting: Using RANSAC for curb fitting is essential because LiDAR data can be noisy and contain outliers, and RANSAC effectively handles these to produce a reliable curb model.
- 1D Kalman Filter for Smoothing Lateral Corrections: Applying a separate Kalman filter for lateral error estimation and correction helps to smooth out the adjustments, preventing jerky movements and ensuring a more stable final trajectory.
Key Results
The experimental results show significant improvements in localization accuracy and robustness:
- Robustness to GNSS Jumps: The proposed UKF method with jump detection successfully maintained accurate trajectories even when random noise was intentionally introduced into GNSS measurements, whereas the method without jump detection deviated significantly.
- Improved Lateral Accuracy: Figures 12-14 demonstrate a substantial reduction in lateral errors after applying the LiDAR-based correction. Figure 14 shows lateral errors reduced from over 1 meter to approximately 0.2 meters over a distance of 140 meters.
- Overall Localization Improvement: The combined approach of fault-tolerant fusion and LiDAR-based lateral correction leads to precise and robust vehicle localization, essential for autonomous driving.
Limitations and Future Work
- Limitations:
- The paper acknowledges that the accuracy of the estimated location is about 1-3 cm or 1 m depending on RTK usage, and even with improvements, lateral offsets can still be an issue without correction.
- The dataset used did not contain encoder data, requiring artificial generation.
- Future Work:
- The authors suggest exploring machine learning and deep learning algorithms for more advanced point cloud processing, specifically for separating curbs from other dynamic obstacles like vehicles and pedestrians.
Summary updated