CFEAR Radarodometry -- Conservative Filtering for Efficient and Accurate Radar Odometry (2105.01457v3)

Abstract: This paper presents the accurate, highly efficient, and learning-free method CFEAR Radarodometry for large-scale radar odometry estimation. By using a filtering technique that keeps the k strongest returns per azimuth and by additionally filtering the radar data in Cartesian space, we are able to compute a sparse set of oriented surface points for efficient and accurate scan matching. Registration is carried out by minimizing a point-to-line metric and robustness to outliers is achieved using a Huber loss. We were able to additionally reduce drift by jointly registering the latest scan to a history of keyframes and found that our odometry method generalizes to different sensor models and datasets without changing a single parameter. We evaluate our method in three widely different environments and demonstrate an improvement over spatially cross-validated state-of-the-art with an overall translation error of 1.76% in a public urban radar odometry benchmark, running at 55Hz merely on a single laptop CPU thread.

• The paper introduces a novel radar odometry method that uses conservative filtering to retain the k strongest radar returns for enhanced noise resistance.
• It employs a two-step approach converting filtered data into oriented surface points in Cartesian space, optimizing scan registration with a robust point-to-line metric.
• Evaluation on benchmark datasets, including the Oxford Radar RobotCar Dataset, demonstrates a 1.76% translation error, affirming its accuracy and adaptability across varied environments.

Overview of CFEAR Radarodometry

The paper introduces CFEAR Radarodometry, a technique for radar odometry focusing on efficiency, accuracy, and robustness without relying on machine learning. This method stands out by employing conservative filtering to retain the strongest $k$ returns per azimuth and further processing in Cartesian space. The resultant sparse set of oriented surface points enhances scan matching efficiency and accuracy. The framework is compatible with various environments and radar sensor models without requiring parameter adjustments, which is validated across three different types of environments.

Methodology

CFEAR Radarodometry is distinguished by its two-step filtering and data representation approach:

1. Filtering: The approach utilizes $k$-strongest filtering of radar returns, which significantly reduces false positives while preserving essential landmark data. This conservative approach surpasses traditional methods like CFAR by maintaining a more accurate and noise-resistant dataset.
2. Oriented Surface Points: The filtered radar data is downsampled into a sparse and efficient representation of oriented surface points. These points are further used in optimizing scan registration through a point-to-line metric, robust against outliers thanks to a Huber loss.

A crucial component of CFEAR is simultaneously registering the latest radar scan against a set of previous keyframes, thereby enhancing robustness and accuracy, particularly in environments with limited or occluded features.

Evaluation and Results

The method was rigorously evaluated in various environments using both public benchmark datasets and collected data from semi-structured and structured environments. In the Oxford Radar RobotCar Dataset, CFEAR consistently delivered superior accuracy, registering a 1.76% translation error, demonstrating a significant improvement over existing state-of-the-art methods under spatial cross-validation conditions.

The qualitative assessments in a vehicle test environment and an underground mine further highlight CFEAR’s adaptability without parameter tuning, although environment-specific adjustments significantly enhanced performance. This flexibility underscores CFEAR's potential for various practical applications, from urban navigation to confined industrial settings.

Implications and Future Directions

CFEAR Radarodometry provides a realistic alternative to learning-based approaches, maintaining accuracy across different datasets without the need for extensive training on large datasets. The straightforward implementation makes it versatile for autonomous navigation tasks, especially in GNSS-denied environments.

While the current method focuses on odometry, its principles could enhance SLAM systems by integrating orientation estimation with mapping efforts. Additionally, exploring different loss functions and incorporating surface uncertainty could further optimize performance.

Overall, CFEAR's contribution lies in its methodological simplicity combined with robust performance, offering a compelling option for radar-based navigation in diverse environmental conditions.