Kidnapped Radar: Topological Radar Localisation using Rotationally-Invariant Metric Learning (2001.09438v1)

Abstract: This paper presents a system for robust, large-scale topological localisation using Frequency-Modulated Continuous-Wave (FMCW) scanning radar. We learn a metric space for embedding polar radar scans using CNN and NetVLAD architectures traditionally applied to the visual domain. However, we tailor the feature extraction for more suitability to the polar nature of radar scan formation using cylindrical convolutions, anti-aliasing blurring, and azimuth-wise max-pooling; all in order to bolster the rotational invariance. The enforced metric space is then used to encode a reference trajectory, serving as a map, which is queried for nearest neighbours (NNs) for recognition of places at run-time. We demonstrate the performance of our topological localisation system over the course of many repeat forays using the largest radar-focused mobile autonomy dataset released to date, totalling 280 km of urban driving, a small portion of which we also use to learn the weights of the modified architecture. As this work represents a novel application for FMCW radar, we analyse the utility of the proposed method via a comprehensive set of metrics which provide insight into the efficacy when used in a realistic system, showing improved performance over the root architecture even in the face of random rotational perturbation.

• The paper introduces a robust method for topological radar localisation using rotationally-invariant deep metric learning on FMCW radar scans.
• It employs adapted deep learning techniques like cylindrical convolutions and azimuth pooling to create radar scan embeddings resilient to rotation, validated on large-scale urban datasets.
• This radar-centric approach provides significant advantages over vision systems in challenging conditions and rotational changes, enhancing autonomous vehicle navigation.

Kidnapped Radar: Topological Radar Localisation using Rotationally-Invariant Metric Learning

This paper introduces a robust approach for large-scale topological localisation using Frequency-Modulated Continuous-Wave (FMCW) radar and presents significant advancements beyond traditional vision-based modalities. The proposed system innovatively integrates deep learning architectures, specifically Convolutional Neural Networks (CNN) and NetVLAD, to learn a metric space tailored for polar radar scan embeddings. This methodological adaptation is achieved through cylindrical convolutions, anti-aliasing blurring, and azimuth-wise max-pooling techniques aimed at achieving rotational invariance, an essential property given the omnidirectional nature of FMCW radar scans.

The architectural improvements ensure radar-based place recognition that is resilient to rotational perturbations, significantly enhancing localisation capabilities. The embedding space learned through this process is used to encode reference trajectories that form a map, which is then queried by nearest neighbor (NN) algorithms to recognize places during autonomous operation.

Empirical validation of the system is conducted using the Oxford Radar RobotCar Dataset, which features extensive urban driving data, comprised of 280 kilometers of radar scans. The results demonstrate improved performance over traditional vision-centric architectures, even when faced with random rotational alterations, thus highlighting the efficacy of radar-centric localisation methods.

Key Numerical Results and Claims

• Improved Precision: The system achieves a 90.49% F1 score in validation, illustrating superior discriminative capabilities in the learned metric space compared to the baseline model.
• Robustness to Rotational Perturbation: Perturbation tests show minimal performance deterioration, indicating the success of architectural modifications aimed at rotational invariance.
• Large-Scale Deployment Feasibility: The method efficiently handles large-scale data without performance degradation, showcasing its applicability to expansive and dynamic urban environments.

Implications

Practical Implications: The introduction of an FMCW radar-focused localisation system offers substantial advancements for autonomous vehicles operating under challenging environmental conditions like rain, fog, or poorly lit scenarios. Traditional vision systems falter in such conditions, whereas radar maintains functionality due to its inherent resilience to atmospheric disturbances.

Theoretical Implications: The work signifies the importance of adapting deep learning models traditionally thresholded to visual data for radar applications, broadening the perspective for metric learning in non-visual domains. The principles demonstrated could foster further research into integrating radar with other sensor modalities, such as LiDAR and UWB, to enhance the robustness and accuracy of localisation systems.

Future Directions: Future work could explore the integration of the presented system with complete Simultaneous Localisation and Mapping (SLAM) frameworks to optimize both place recognition and metric pose estimation processes. Additionally, techniques for embedding space discretization could adopt approximate NN search methods to improve computational efficiency for real-time applications.

In conclusion, the paper incisively expands the utilisation of radar technology within the sphere of autonomous navigation, vindicated by practical deployment outcomes and potential theoretical contributions to metric learning and localisation strategies. This progression forms a basis for broader adoption and evolution of radar-based systems in vehicular autonomy.