Off the Radar: Uncertainty-Aware Radar Place Recognition with Introspective Querying and Map Maintenance

Jianhao Yuan, Paul Newman, Matthew Gadd

TL;DR

Researchers from the University of Oxford have developed a new framework for radar-based place recognition that doesn't just guess where a robot is—it knows when it’s likely to be wrong. By leveraging variational contrastive learning, the system identifies "noisy" radar scans to improve map quality over time and introspectively reject unreliable localization queries, significantly outperforming current state-of-the-art methods.

What Problem Does This Paper Address?

Autonomous vehicles often rely on cameras or LiDAR for localization (place recognition). However, these sensors struggle in "adversarial" conditions like heavy rain, fog, or direct sunlight. Frequency-Modulated Continuous-Wave (FMCW) radar is a robust alternative because it "sees" through these conditions.

The catch? Radar data is messy. It’s filled with artifacts like speckle noise, saturation, and ghosting. Most existing radar localization algorithms treat every scan as equally valid, which can lead to confident but incorrect "false positive" matches—a nightmare for safety-critical systems. The authors address the lack of calibrated uncertainty estimation in radar place recognition: the ability for a system to quantify its own confidence in real-time.

Methods

The authors propose a unified system divided into offline learning and online deployment stages:

  1. Variational Contrastive Learning (VCL): In the offline stage, they use a Variational Autoencoder (VAE) structure to disentangle radar embeddings. They split the representation into two parts: a semantic invariant part ($Z_I$) for identifying the location and a noise-induced variant part ($Z_V$) that captures uncertainty.
  2. Four-Way Loss Function: To train this, they use a combination of:
  3. Invariant Contrastive Loss: Ensures similar places have similar semantic embeddings.
  4. Variant Contrastive Loss: Helps structure the uncertainty space.
  5. KL Divergence: Regularizes the uncertainty against a prior distribution.
  6. Reconstruction Loss: Ensures the embedding retains enough information to "rebuild" the original feature map.
  7. Continual Map Maintenance: During online operation, the system doesn't just use a static map. If it traverses a known route and captures a "cleaner" (lower uncertainty) scan than what's currently in the map dictionary, it replaces the old scan.
  8. Introspective Query Rejection: When the robot asks "Where am I?", the system first checks the uncertainty of the current scan. If the uncertainty exceeds a static threshold, the system rejects the query entirely rather than risking a wrong guess.

Datasets

The study utilizes two massive, well-known benchmarks in the robotics community:

  • Oxford Radar RobotCar: Contains 300 km of driving data through Oxford city center across various traversals.
  • MulRan Dataset: Provides 120 km of data from four distinct locations in Daejeon, South Korea, known for being particularly challenging due to sparse environments.

Key Insights and Ideas

  • Aleatoric vs. Epistemic Uncertainty: The paper argues that for radar, aleatoric uncertainty (noise inherent in the data, like motion blur) is more important than epistemic uncertainty (noise from the model's lack of knowledge). Their VAE approach targets this data-level noise.
  • The "Upgrade" Strategy: Most localization systems treat the map as a "source of truth." This paper treats the map as a living document that can be improved by swapping out "blurry" or "saturated" scans for high-quality ones over multiple traversals.
  • Static vs. Dynamic Thresholding: Unlike previous methods (like STUN) that require a whole batch of scans to determine what is "uncertain," this method uses a static threshold. This is crucial for real-time robotics where decisions must be made frame-by-frame.

Key Results

  • State-of-the-Art Performance: On the Oxford dataset, their method achieved a 93.67% Recall@1, outperforming the previous best (STUN) by over 4%.
  • Reliable Rejection: In both datasets, as the system was allowed to reject more "uncertain" queries, the accuracy of the remaining queries increased consistently. Competing methods like MC Dropout and STUN often saw performance decrease or fluctuate when trying to reject queries on the MulRan dataset.
  • Map Quality: The map maintenance algorithm alone boosted Recall@1 from 90.46% to 93.67%, proving that being selective about map data pays off.

Limitations and Future Work

  • GPS Dependency for Maintenance: Currently, the "Map Maintenance" step uses GPS to associate new scans with map locations offline. While the authors note this could be replaced by radar-only SLAM, it remains a dependency in the current pipeline.
  • Sparse Environments: The system still finds sparse, open areas (common in the MulRan dataset) difficult, as these environments produce radar scans that look identical regardless of the specific location.
  • Future Directions: The authors suggest exploring more complex "spinning" strategies for data augmentation and integrating the uncertainty measure into a full SLAM (Simultaneous Localization and Mapping) backend to improve global consistency.

Summary updated

Abstract

Localisation with Frequency-Modulated Continuous-Wave (FMCW) radar has gained increasing interest due to its inherent resistance to challenging environments. However, complex artefacts of the radar measurement process require appropriate uncertainty estimation - to ensure the safe and reliable application of this promising sensor modality. In this work, we propose a multi-session map management system which constructs the “best” maps for further localisation based on learned variance properties in an embedding space. Using the same variance properties, we also propose a new way to introspectively reject localisation queries that are likely to be incorrect. For this, we apply robust noise-aware metric learning, which both leverages the short-timescale variability of radar data along a driven path (for data augmentation) and predicts the downstream uncertainty in metric-space-based place recognition. We prove the effectiveness of our method over extensive cross-validated tests of the Oxford Radar RobotCar and MulRan dataset. In this, we outperform the current state-of-the-art in radar place recognition and other uncertainty-aware methods when using only single nearest-neighbour queries. We also show consistent performance increases when rejecting queries based on uncertainty over a difficult test environment, which we did not observe for a competing uncertainty-aware place recognition system.
BibTeX
@inproceedings{yuan2023off,
  author = {Jianhao Yuan and Paul Newman and Matthew Gadd},
  title = {{Off the Radar: Uncertainty-Aware Radar Place Recognition with Introspective Querying and Map Maintenance}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2023},
  url = {https://openreview.net/forum?id=v2npGWfijV},
}

Loading related content...

Figure