AirVO: An Illumination-Robust Point-Line Visual Odometry (2212.07595v3)

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.

• The paper introduces a hybrid VO system that fuses CNN and GNN-based feature extraction with an extended line processing pipeline to achieve robust performance under varying illumination.
• It demonstrates over five-fold speed improvements by running at 15Hz on low-power embedded systems while consistently reducing translational errors across challenging datasets.
• The innovative approach bridges advanced learning methods and traditional optimization techniques to enhance real-time visual odometry in difficult lighting conditions.

Overview of "AirVO: An Illumination-Robust Point-Line Visual Odometry"

This paper presents AirVO, a novel illumination-robust Visual Odometry (VO) system integrating learning-based corner point algorithms with an extended line feature algorithm. The system leverages both Convolutional Neural Networks (CNNs) and Graph Neural Networks (GNNs) for detecting and matching influential corner points, thereby addressing challenges posed by dynamic illumination variations. Particularly, the capability to operate in real-time on lightweight, low-power embedded platforms is a significant contribution, achieved through optimized acceleration of both CNN and GNN components.

Methodology

The core innovation of AirVO lies in its hybrid approach which combines learning-based feature extraction with traditional optimization techniques. The system introduces methods for robust point matching and an efficient line processing pipeline, facilitating improved accuracy in difficult lighting conditions.

1. Feature Extraction and Matching: Using SuperPoint for feature extraction and SuperGlue for matching, AirVO capitalizes on efficient and reliable point tracking. The integration of learning-based methods ensures robustness beyond what handcrafted methods can achieve, especially under extreme lighting conditions.
2. Line Processing Pipeline: The line processing involves associating 2D points with detected lines and using the matching results from these associated points to triangulate and match lines effectively. This robust methodology challenges existing limitations in line detection, making the system resilient against unstable conditions that typically impact line tracking and matching.
3. System Architecture and Optimization: AirVO utilizes a multi-threaded pipeline to optimize CPU and GPU resource usage. The learning-based components are crucially accelerated, providing more than five-fold improvement in performance over similar systems, allowing the VO to run efficiently at 15Hz on embedded devices.

Experimental Validation

AirVO's capabilities were evaluated on multiple datasets, specifically targeting environments with variable lighting. The results demonstrated its superiority over existing state-of-the-art VO and Visual Inertial Odometry (VIO) systems, particularly in terms of robustness and accuracy, with significant benchmarks set by metrics such as Root Mean Square Error (RMSE). Notably, the system consistently showed lower translational errors across diverse challenging scenarios.

Implications and Future Work

AirVO's development marks a significant step forward in making visual odometry systems more applicable in real-world scenarios where lighting conditions can fluctuate unpredictably. The implications of this research are manifold:

• Practical Applications: The ability to run on low-power devices extends the applicability to a range of platforms, including drones and mobile robots, where power efficiency is critical.
• Robust Feature Tracking: The resilience against dynamic illumination opens paths for enhanced long-term navigation and mapping within unstructured environments.
• Foundation for SLAM Systems: Future developments might extend AirVO to a comprehensive SLAM system, incorporating loop closure and re-localization for detailed mapping and localization tasks.

Conclusion

AirVO effectively bridges the gap between learning-based feature extraction and traditional computational efficiency to enhance visual odometry systems. This hybrid methodology not only positions AirVO as a leading edge solution in its current form but also sets a precedent for future research and development in robust, efficient, and reliable visual odometry and SLAM systems applicable to real-world dynamic environments. The open-source release of AirVO further underscores its potential as a catalyst for ongoing innovation within the robotics and computer vision communities.