基于双目相机的频域特征快速回环检测算法研究

A Fast loop closure detection for Frequency-Domain Features Using Stereo Cameras

摘要:

针对传统回环检测算法在定位精度和计算效率方面的不足，以及基于Transformer的方法计算开销较大的问题，本文提出了一种基于双目视觉和频域特征的高效回环检测方法。该方法利用轻量级特征提取模块，从双目图像中提取关键特征，并构建全局描述子，以提高候选关键帧的筛选效率。此外，本文引入快速傅里叶变换构建频域描述子，从频域角度对局部特征进行重新表征，并利用频域特征匹配对候选关键帧进行重排序，同时优化底层计算策略以加速图像匹配过程。实验结果表明，该算法在New College数据集上召回率最优，在Oxford和KITTI数据集上性能超过80%的算法，且图像匹配速度提高了1-2倍。在确保较高定位精度的同时，该方法有效降低计算成本，显著提升回环检测的鲁棒性和效率。

Abstract:

To address the limitations of traditional loop closure detection in terms of localization accuracy and computational efficiency, as well as the high computational cost of Transformer-based methods, this paper proposes an efficient loop detection approach utilizing stereo vision and frequency-domain features. A lightweight feature extraction module is employed to extract key features from stereo images and construct global descriptors, enhancing the efficiency of candidate keyframes selection. The Fast Fourier Transform is introduced to build frequency-domain descriptors, providing a spectral representation of local features. Frequency-domain feature matching is then applied to refine the ranking of candidate keyframes. Meanwhile, a tailored low-level computational strategy accelerates the image matching process. Experimental results show that our method achieves the highest recall on the New College dataset, outperforms 80% of existing algorithms on the Oxford and KITTI datasets, and accelerates image matching by a factor of 1-2. It ensures high localization accuracy, reduces computational costs, and significantly improves the robustness and efficiency of loop detection.