Abstract
Accurate estimation of the ground plane normal is critical for autonomous driving tasks such as Bird’s-Eye-View (BEV) perception and 3D object detection. Traditional methods often rely on static extrinsic calibration, which fails to account for dynamic vehicle pitch and roll during acceleration, braking, or travel over uneven surfaces. This paper presents a real-time, sensor-agnostic approach that estimates the dynamic ground plane normal using only ego-motion (odometry) data. By employing an Invariant Extended Kalman Filter (IEKF), the proposed method achieves state-of-the-art accuracy with minimal computational overhead.
1. Introduction
In the field of autonomous driving, projecting 2D image data into a 3D Bird’s-Eye-View (BEV) space is a foundational step for lane detection, path planning, and object tracking. Most existing systems assume a "static extrinsic calibration," where the camera's orientation relative to the road surface is considered fixed. However, real-world driving conditions involve constant physical dynamics; even a subtle 1° oscillation in vehicle pitch can lead to significant projection errors and "wobbling" in the BEV representation.
While high-end sensors like LiDAR or stereo cameras can measure the ground plane directly, they are often cost-prohibitive or computationally expensive. Conversely, monocular deep learning methods frequently struggle to capture subtle physical dynamics and suffer from high latency. This research proposes a middle ground: a highly efficient, geometry-based filter that uses the vehicle's own motion as a proxy for the ground plane orientation.
2. Methodology
The core of the proposed system is a pipeline that decouples ground plane estimation from visual perception, focusing instead on the relationship between vehicle dynamics and the road surface.
2.1 Ego-Motion as Input
The system is designed to be sensor-agnostic, accepting ego-motion data from various sources, including monocular SLAM, Structure from Motion (SfM), or Inertial Measurement Units (IMU). The fundamental insight is that the vehicle’s physical oscillations directly reflect the ground plane's relationship to the onboard sensors.
2.2 Invariant Extended Kalman Filter (IEKF)
To process this motion data, the researchers utilize an Invariant Extended Kalman Filter (IEKF). Unlike standard EKFs, the IEKF is formulated on Lie groups (specifically SO(3) for rotations). This mathematical framework is more robust for tracking orientations, as it avoids the singularities and linearization errors that typically plague traditional filters when dealing with rotational geometry.
2.3 Frequency Splitting and State Model
The filter distinguishes between two primary components of vehicle motion:
- Low-frequency components: Actual changes in road topography, such as hills or gradients.
- High-frequency components: The "wobble" or momentary oscillations of the vehicle body caused by suspension response.
Surprisingly, the authors found that a simple zero-order (constant) process model within the IEKF was sufficient to track these complex dynamics accurately in real-time, provided the filter was correctly tuned to these frequency components.
3. Experimental Evaluation
The method was evaluated using two prominent autonomous driving benchmarks:
- KITTI Odometry Dataset: Used for primary quantitative analysis. Ground truth normal vectors were generated by applying RANSAC plane fitting to LiDAR point clouds.
- nuScenes Dataset: Used to test the generalization and robustness of the approach across diverse urban environments.
4. Results
4.1 Accuracy
On the KITTI dataset, the proposed method achieved a vector error of only 0.39°. This represents a significant improvement over previous benchmarks:
- GroundNet (Deep Learning): 0.70° error.
- Traditional Homography: 4.10° error.
The research quantifies that a standard 4-wheel sedan typically experiences pitch oscillations of approximately 1° even on relatively flat roads, making the 0.39° accuracy threshold vital for stable perception.
4.2 Computational Efficiency
The IEKF component is exceptionally lightweight, requiring less than 1 millisecond of execution time. When integrated with IMU-based odometry, the total system latency is approximately 3ms per frame. This is orders of magnitude faster than CNN-based approaches, which typically range from 130ms to 920ms, making this method ideal for deployment on low-power embedded hardware.
4.3 Qualitative Impact
Visualizations of Inverse Perspective Mapping (IPM) demonstrate the practical benefits of this method. While static calibration results in "fishtailing" or unstable road edges in the BEV projection, the proposed dynamic estimation keeps road features perfectly parallel and stable, even during aggressive vehicle maneuvers.
5. Discussion and Limitations
While highly effective, the method relies on the physical connection between a wheeled chassis and the ground. Consequently, it may not be directly applicable to non-wheeled platforms, such as legged robots or drones. Additionally, the model assumes a "local flatness" of the road surface; while sufficient for most paved environments, accuracy may decrease on highly irregular or jagged off-road terrain. Finally, during extreme lateral acceleration (sharp turns), the vehicle's dynamics may exceed the complexity of the current zero-order model, suggesting an area for future refinement using higher-order states.
6. Conclusion
This paper demonstrates that accurate ground plane estimation does not require complex deep learning models or expensive depth sensors. By leveraging the geometric properties of ego-motion through an Invariant Extended Kalman Filter, the researchers provide a robust, real-time solution that significantly enhances the stability of BEV perception systems in autonomous vehicles.
Summary updated