Under the Radar: Learning to Predict Robust Keypoints for Odometry Estimation and Metric Localisation in Radar (2001.10789v3)

Abstract: This paper presents a self-supervised framework for learning to detect robust keypoints for odometry estimation and metric localisation in radar. By embedding a differentiable point-based motion estimator inside our architecture, we learn keypoint locations, scores and descriptors from localisation error alone. This approach avoids imposing any assumption on what makes a robust keypoint and crucially allows them to be optimised for our application. Furthermore the architecture is sensor agnostic and can be applied to most modalities. We run experiments on 280km of real world driving from the Oxford Radar RobotCar Dataset and improve on the state-of-the-art in point-based radar odometry, reducing errors by up to 45% whilst running an order of magnitude faster, simultaneously solving metric loop closures. Combining these outputs, we provide a framework capable of full mapping and localisation with radar in urban environments.

• The paper demonstrates that a self-supervised framework for learning robust keypoints reduces radar odometry errors by up to 45%.
• It employs a differentiable pose estimator with SVD to optimize keypoint predictions directly from localization errors.
• The sensor-agnostic design and U-Net based architecture enhance real-time radar localization and enable effective loop closure detection.

Learning to Predict Robust Keypoints for Radar-Based Localization

The paper "Under the Radar: Learning to Predict Robust Keypoints for Odometry Estimation and Metric Localisation in Radar" discusses a self-supervised framework aimed at improving the accuracy and efficiency of odometry estimation and localization specifically using radar data. The approach focuses on learning to predict robust keypoints, a paradigm shift in which the optimization of these keypoints is driven explicitly by the application needs rather than predefined characteristics. This has particular significance given the complexity of radar data, which is often fraught with challenges such as noise artifacts and limited precision compared to more conventional modalities like vision and LiDAR.

Methodology and Design

This work leverages a novel self-supervised architecture that incorporates a differentiable point-based motion estimator, enabling the training process to utilize localization errors directly for supervision. Keypoint locations, scores, and descriptors are optimized based on their influence on localization accuracy rather than relying on hand-crafted heuristics. A notable feature of the proposed method is its sensor-agnostic configuration, extending its applicability beyond radar to potentially encompass a variety of sensor types.

The formulation includes several distinctive components:

• Keypoint Detection: Utilizes a U-Net style convolutional encoder with multiple decoders to predict keypoint parameters at high resolution. The points predicted are associated with scores that estimate their significance for odometry tasks and descriptors to assist in identifying points across different instances of the same scene.
• Differentiable Pose Estimation: The algorithm matches keypoints from sequential radar scans and includes a pose estimator that utilizes singular value decomposition (SVD) to compute optimal transformations. This differentiable process allows seamless backpropagation of errors to the keypoint prediction stage.
• Metric Localisation: Extracts location-specific embeddings from the radar data, allowing for the detection of topological loop closures and providing the foundation for complete metric localization.

Empirical Evaluation

The experiments were primarily conducted using the Oxford Radar RobotCar Dataset, covering approximately 280 kilometers of urban driving environments. The results demonstrated a notable improvement—up to 45% reduction in errors—over existing point-based radar odometry techniques, with the proposed method operating at a computational speed significantly faster than prior benchmarks. Additionally, the framework successfully identified and resolved metric loop closures, indicating its viability as a comprehensive mapping and localization solution.

Implications and Future Directions

The implications of this research are multifaceted, offering both theoretical and practical advancements. Theoretically, it highlights the potential of self-supervised learning in scenarios where direct supervision is impractical, promoting an adaptable architecture that prioritizes specific task outcomes. Practically, this innovation enhances the potential of radar in autonomous navigation systems, complementing its benefits like robustness to adverse lighting and weather conditions with improved localization precision.

Looking ahead, the research opens pathways for applying this approach to other sensors, potentially incorporating them into a multi-modal solution that capitalizes on the strengths of each modality. The adaptability of the keypoint detection mechanism hints at broader applicability in various real-time systems. Further exploration could involve integrating this system with machine learning models for tasks extending beyond localization, such as dynamic object tracking or semantic mapping, thus enriching the suite of competencies available for autonomous vehicles in increasingly complex environments.