ORB-SLAM3AB: Augmenting ORB-SLAM3 to Counteract Bumps with Optical Flow Inter-frame Matching

Yangrui Dong,Weisheng Gong,Qingyong Li,Kaijie Su,Chen He,Z. Jane Wang

2024-11-27

Abstract:This paper proposes an enhancement to the ORB-SLAM3 algorithm, tailored for applications on rugged road surfaces. Our improved algorithm adeptly combines feature point matching with optical flow methods, capitalizing on the high robustness of optical flow in complex terrains and the high precision of feature points on smooth surfaces. By refining the inter-frame matching logic of ORB-SLAM3, we have addressed the issue of frame matching loss on uneven roads. To prevent a decrease in accuracy, an adaptive matching mechanism has been incorporated, which increases the reliance on optical flow points during periods of high vibration, thereby effectively maintaining SLAM precision. Furthermore, due to the scarcity of multi-sensor datasets suitable for environments with bumpy roads or speed bumps, we have collected LiDAR and camera data from such settings. Our enhanced algorithm, ORB-SLAM3AB, was then benchmarked against several advanced open-source SLAM algorithms that rely solely on laser or visual data. Through the analysis of Absolute Trajectory Error (ATE) and Relative Pose Error (RPE) metrics, our results demonstrate that ORB-SLAM3AB achieves superior robustness and accuracy on rugged road surfaces.

Robotics

What problem does this paper attempt to address?

This paper attempts to solve the problem of performance degradation of visual SLAM (Simultaneous Localization and Mapping) systems under rough road conditions. Specifically, existing visual SLAM algorithms such as ORB - SLAM3 will experience a significant decline in the accuracy of localization and mapping due to the loss of feature point matching when encountering bumps or sharp turns. To solve this problem, the paper proposes an enhanced version of the ORB - SLAM3 algorithm - ORB - SLAM3AB, which improves robustness and accuracy in complex terrains by combining the optical flow method and the feature point matching method.

### Main contributions:
1. **Enhanced inter - frame matching**: By introducing optical flow technology, the inter - frame matching strategy of ORB - SLAM3 is improved, overcoming the limitations of feature - point - based matching, and especially enhancing the system's robustness in bumpy and sharp - turn scenarios.
2. **Comprehensive comparative analysis**: A comprehensive comparative analysis of the improved algorithm was carried out, including comparisons with laser SLAM and pure monocular visual SLAM methods. By plotting trajectory maps and calculating the Absolute Trajectory Error (ATE) and Relative Pose Error (RPE), the results show that ORB - SLAM3AB has higher accuracy and robustness on bumpy roads.
3. **Dataset collection**: A large amount of data was collected on low - speed robots and high - speed vehicles, covering bumpy sections under various weather, lighting, and road conditions, providing a valuable resource for evaluating the performance of SLAM under these challenging conditions.

### Technical details:
- **Image pre - processing**: The input key - frame images and current - frame images are converted into grayscale images after color channel inspection, and Gaussian filtering is applied to reduce noise and improve the quality of feature extraction.
- **Feature point extraction**: Key points and descriptors are extracted from the grayscale images, and the parameters of the feature point extractor (such as the number of detection points, scale factor, and number of pyramid levels) are dynamically adjusted according to the detection strategy.
- **Dynamic optical - flow feature - point detection and matching**: Initially, the number of optical - flow points is set to half of the number of feature points. If it is found that the number of matching feature points in the current frame is insufficient during the subsequent matching process, the number of optical - flow feature points is dynamically increased to ensure that inter - frame matching can still be carried out in bumpy situations. When the number of matching points is sufficient, the number of optical - flow feature points is reduced to improve the accuracy of SLAM.
- **Rotation consistency check**: A rotation consistency check is performed on the matched feature - point pairs, and the Histogram of Oriented Gradients (HOG) method is used to count the rotation angles of the feature points to remove potential false matches.

### Experimental results:
- Tests were carried out under various conditions such as low - speed sunny days and nights, and high - speed snowy days. The results show that ORB - SLAM3AB performs significantly better than other SLAM algorithms on bumpy sections, especially in high - vibration and sharp - turn scenarios.
- Through the comparison of ATE and RPE indicators, ORB - SLAM3AB shows higher accuracy and robustness under various bumpy conditions.

### Conclusion:
The paper significantly improves the performance of ORB - SLAM3 on bumpy roads by combining the optical flow technology and the feature point matching method. Future research can further optimize the algorithm to improve the robustness of bumpy sections while maintaining the accuracy on smooth roads. In addition, exploring sensor fusion (such as adding LiDAR data) to enhance the algorithm's robustness in severely bumpy situations is also a research direction worthy of exploration.