Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
126 tokens/sec
GPT-4o
47 tokens/sec
Gemini 2.5 Pro Pro
43 tokens/sec
o3 Pro
4 tokens/sec
GPT-4.1 Pro
47 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Masking by Moving: Learning Distraction-Free Radar Odometry from Pose Information (1909.03752v4)

Published 9 Sep 2019 in cs.CV, cs.LG, and cs.RO

Abstract: This paper presents an end-to-end radar odometry system which delivers robust, real-time pose estimates based on a learned embedding space free of sensing artefacts and distractor objects. The system deploys a fully differentiable, correlation-based radar matching approach. This provides the same level of interpretability as established scan-matching methods and allows for a principled derivation of uncertainty estimates. The system is trained in a (self-)supervised way using only previously obtained pose information as a training signal. Using 280km of urban driving data, we demonstrate that our approach outperforms the previous state-of-the-art in radar odometry by reducing errors by up 68% whilst running an order of magnitude faster.

Citations (78)

Summary

  • The paper introduces a deep learning framework for distraction-free radar odometry using a fully differentiable, correlation-based scan matching approach.
  • It employs self-supervised training with pose information, eliminating the need for external ground-truth data.
  • Experimental results over 280 km of urban driving data show up to 68% error reduction and an order of magnitude faster processing than state-of-the-art methods.

End-to-End Radar Odometry System for Autonomous Driving

This paper introduces an advanced radar odometry system specifically designed to provide robust and real-time pose estimates. The focal point of this research is the development of a radar-based odometry system grounded in a learned embedding space, which is inherently free from sensing artefacts and distractor objects. The system's innovation lies in employing a fully differentiable, correlation-based radar matching approach, ensuring both interpretability and uncertainty quantification comparable to traditional scan-matching methods.

Methodology

The proposed system utilizes deep correlative scan matching integrated with learned feature embeddings to achieve accurate radar odometry. By operating within an end-to-end learning framework, the system is trained using a self-supervised methodology, relying solely on previously obtained pose information. This approach circumvents the need for ground-truth data derived from external sources, thus streamlining the training process.

Experimental Validation

The experimental evaluation employs 280 km of urban driving data, demonstrating the system's capability to reduce errors by up to 68% compared to existing state-of-the-art radar odometry techniques. Notably, this performance improvement is achieved while maintaining computational efficiency, running an order of magnitude faster than its predecessors.

Results and Discussion

The results section highlights the significant advancement in translational and rotational accuracy. Through detailed analysis, the paper quantitatively illustrates the performance enhancement over current benchmark radar odometry systems, made evident by the substantial reduction in error margins and processing time.

Theoretical and Practical Implications

Theoretically, this research establishes a method for leveraging deep learning approaches to enhance radar-based localization systems, paying particular attention to the challenges posed by urban environments. Practically, the improvements in odometry precision and real-time processing capabilities present significant advancements for the deployment of autonomous vehicles in complex and dynamic settings.

Future Directions

While this paper showcases marked improvements in radar odometry, future research could focus on extending the applicability of this approach to other sensor modalities, enhancing robustness under extreme weather conditions, and further optimizing the training framework to accommodate broader datasets. Additionally, integrating this radar odometry system with high-level navigation frameworks could potentially unlock new capabilities in autonomous systems deployment.

Conclusion

This paper makes a substantial contribution to the field of autonomous driving by presenting advancements in radar odometry. Leveraging a learned embedding space and a fully differentiable, correlation-based matching approach, it demonstrates significant improvements in accuracy and processing speed, presenting a viable path forward for real-time autonomous navigation systems. The results reinforce the viability of radar systems in complementing vision-based approaches, paving the way for more comprehensive and resilient autonomous driving solutions.

Youtube Logo Streamline Icon: https://streamlinehq.com