Radar Odometry Combining Probabilistic Estimation and Unsupervised Feature Learning (2105.14152v3)

Abstract: This paper presents a radar odometry method that combines probabilistic trajectory estimation and deep learned features without needing groundtruth pose information. The feature network is trained unsupervised, using only the on-board radar data. With its theoretical foundation based on a data likelihood objective, our method leverages a deep network for processing rich radar data, and a non-differentiable classic estimator for probabilistic inference. We provide extensive experimental results on both the publicly available Oxford Radar RobotCar Dataset and an additional 100 km of driving collected in an urban setting. Our sliding-window implementation of radar odometry outperforms most hand-crafted methods and approaches the current state of the art without requiring a groundtruth trajectory for training. We also demonstrate the effectiveness of radar odometry under adverse weather conditions. Code for this project can be found at: https://github.com/utiasASRL/hero_radar_odometry

• The paper presents an unsupervised method that extracts radar features without requiring groundtruth, reducing data collection efforts.
• It integrates classical probabilistic estimators with deep learning to effectively manage uncertainties in real-world settings.
• Extensive experiments on challenging urban datasets demonstrate robust trajectory tracking even under adverse weather conditions.

Overview of Unsupervised Radar Odometry

The paper "Radar Odometry Combining Probabilistic Estimation and Unsupervised Feature Learning" presents a novel approach to radar odometry that eschews reliance on supervised learning with groundtruth trajectory data. Instead, the authors propose a method that leverages on-board radar data to train a feature network in an unsupervised manner, with a strong emphasis on probabilistic trajectory estimation.

Key Contributions

1. Unsupervised Feature Learning: The proposed radar odometry technique employs unsupervised learning to directly extract features from radar data, eliminating the dependency on accurate groundtruth pose information. This aspect is particularly notable for applications where obtaining precise groundtruth is difficult or expensive.
2. Probabilistic Estimation and Deep Learning: The method combines classical non-differentiable estimators with deep learning techniques, which allows for enhanced processing of rich radar data. The classical estimator enables probabilistic inference, which is crucial for handling uncertainties in real-world scenarios.
3. Experimental Validation: The effectiveness of this approach is substantiated through extensive experiments on the Oxford Radar RobotCar Dataset along with a custom dataset encompassing 100 km of driving in urban conditions. The results demonstrate that their sliding-window radar odometry approach outperforms several hand-crafted methods and approaches the performance of current state-of-the-art techniques without needing any groundtruth for training.
4. Robustness to Weather Conditions: The method is shown to be particularly robust under adverse weather conditions, which is an essential characteristic for outdoor navigation systems relying on radar data.

Methodology

The authors utilize the Exactly Sparse Gaussian Variational Inference (ESGVI) for parameter learning, which supports nonlinear batch state estimation. This framework is adept at jointly optimizing model parameters and state with a data likelihood objective, allowing the network to learn radar features conducive to odometry.

The network architecture employed is designed to output detector scores for keypoint detection, weight scores predicting keypoint uncertainty, and descriptors for matching. A distinctive feature of the method is its modular design, which makes it adaptable for integration with additional sensor data and constraints, potentially enhancing its estimation performance.

Implications and Future Directions

The introduction of an unsupervised learning paradigm for radar odometry presents significant implications for the deployment of autonomous systems in GPS-denied or difficult-to-map environments. The ability to train on large datasets without groundtruth can dramatically reduce dataset collection costs and efforts.

This research invites further exploration and extension in several dimensions:

• Integration with Multi-modal Sensor Data: Future work can focus on integrating this radar-based approach with data from other sensors such as LiDAR and cameras, thereby enriching the feature set and potentially improving odometry accuracy.
• Enhanced Outlier Rejection: Although the current method integrates robust outlier rejection schemes, ongoing refinement could increase system reliability, especially under challenging environmental conditions.
• Real-time Implementation: While the current implementation is real-time capable, increased computational efficiency and potential hardware acceleration could extend its applicability to more resource-constrained platforms.

In summary, this paper significantly advances the field of unsupervised radar odometry by combining the strengths of deep learning and probabilistic estimation, paving the way for robust, scalable, and less resource-intensive solutions in autonomous navigation.