AirVO: An Illumination-Robust Point-Line Visual Odometry

Kuan Xu, Yuefan Hao, Chen Wang, Lihua Xie

Abstract

This paper proposes an illumination-robust visual odometry (VO) system that incorporates both accelerated learning-based corner point algorithms and an extended line feature algorithm. To be robust to dynamic illumination, the proposed system employs the convolutional neural network (CNN) and graph neural network (GNN) to detect and match reliable and informative corner points. Then point feature matching results and the distribution of point and line features are utilized to match and triangulate lines. By accelerating CNN and GNN parts and optimizing the pipeline, the proposed system is able to run in real-time on low-power embedded platforms. The proposed VO was evaluated on several datasets with varying illumination conditions, and the results show that it outperforms other state-of-the-art VO systems in terms of accuracy and robustness.

Introduction

Visual Odometry (VO) is a critical component for autonomous robot navigation. However, traditional VO systems often struggle in environments with extreme or dynamic lighting conditions. Most conventional methods rely on the "brightness consistency" assumption, which posits that the appearance of a feature point remains constant across consecutive frames. In real-world scenarios—such as tunnels, mines, or buildings with flickering lights—this assumption frequently fails, leading to tracking loss in systems using handcrafted features like ORB or FAST.

While deep-learning-based features (e.g., SuperPoint) offer significantly better robustness to lighting changes, their high computational cost typically precludes real-time execution on embedded hardware. Furthermore, while line features provide valuable geometric constraints, they are difficult to match reliably when lighting is unstable. AirVO addresses these challenges by combining accelerated deep-learning point tracking with a novel, descriptor-free line matching strategy.

Methodology

AirVO utilizes a hybrid approach that leverages the strengths of both learning-based and geometric methods. The system architecture is divided into a multi-threaded frontend and backend to maximize efficiency.

The AirVO system architecture, highlighting the parallelized frontend (GPU/CPU) and backend threads for point-line visual odometry.

Accelerated Deep Learning

The frontend employs SuperPoint for feature detection and SuperGlue for feature matching. To achieve real-time performance on embedded platforms like the Nvidia Jetson AGX Xavier, the authors optimized the pipeline by converting computations from 32-bit to 16-bit floating-point (FP16). This optimization, combined with pipeline refinements, resulted in a 5x speedup over standard implementations with negligible loss in precision.

Descriptor-Free Line Matching

To overcome the instability of traditional line descriptors in varying light, AirVO introduces a descriptor-free matching method. Lines are associated with nearby robustly tracked points. If the points associated with a line are successfully matched across frames, the line itself is considered a match. Additionally, the system improves upon the standard Line Segment Detector (LSD) by merging fragmented segments into longer, more stable lines.

Hybrid Optimization

The backend performs local bundle adjustment using a unified mathematical framework that incorporates both point and line constraints. Lines are represented using Plücker coordinates, allowing for efficient geometric optimization within the pose estimation process.

Experimental Results

The system was evaluated on the OIVIO (tunnels/mines) and UMA-VI (challenging indoor illumination) datasets.

Trajectory comparison and visual snapshots under dynamic lighting. AirVO demonstrates superior robustness during rapid motion and illumination transitions compared to OKVIS and PL-SLAM.

Accuracy and Robustness

On the OIVIO dataset, AirVO achieved the lowest translational error in 7 out of 9 sequences, outperforming VINS-Fusion and ORB-SLAM2. In the UMA-VI "lights off" tests, where many baseline systems failed entirely, AirVO maintained tracking with an average error roughly 50% lower than its closest competitor, OKVIS. Ablation studies revealed that the inclusion of line features reduced error by approximately 58% compared to a point-only configuration.

Real-Time Performance

AirVO achieves a processing rate of 15Hz on an Nvidia Jetson AGX Xavier and up to 40Hz on a standard laptop. This demonstrates that deep-learning-based VO can be successfully deployed on edge devices for real-world robotics applications.

Conclusion

AirVO presents a robust solution for visual odometry in challenging lighting environments. By integrating accelerated deep-learning point features with a novel line-tracking pipeline, the system achieves state-of-the-art accuracy and real-time performance on embedded hardware. Future work will focus on extending AirVO into a full SLAM system with loop closure and global mapping capabilities.

Summary updated

Abstract

This paper proposes an illumination-robust visual odometry (VO) system that incorporates both accelerated learning-based corner point algorithms and an extended line feature algorithm. To be robust to dynamic illumination, the proposed system employs the convolutional neural network (CNN) and graph neural network (GNN) to detect and match reliable and informative corner points. Then point feature matching results and the distribution of point and line features are utilized to match and triangulate lines. By accelerating CNN and GNN parts and optimizing the pipeline, the proposed system is able to run in real-time on low-power embedded platforms. The proposed VO was evaluated on several datasets with varying illumination conditions, and the results show that it outperforms other state-of-the-art VO systems in terms of accuracy and robustness. The open-source nature of the proposed system allows for easy implementation and customization by the research community, enabling further development and improvement of VO for various applications.
BibTeX
@inproceedings{xu2023airvo,
  author = {Kuan Xu and Yuefan Hao and Chen Wang and Lihua Xie},
  title = {{AirVO: An Illumination-Robust Point-Line Visual Odometry}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2023},
  doi = {10.1109/iros55552.2023.10341914},
}
Figure