Huang Yizhe, Huang Bin, Zhang Zhifu, Shi Yuanyuan, Yuan Yizhao, Sun Jinfeng
Hubei Key Laboratory of Modern Manufacturing Quality Engineering, School of Mechanical Engineering, Hubei University of Technology, Wuhan 430068, China.
State Key Laboratory of Digital Manufacturing Equipment and Technology, Huazhong University of Science and Technology, Wuhan 430074, China.
Sensors (Basel). 2023 Oct 22;23(20):8632. doi: 10.3390/s23208632.
Variations with respect to perspective, lighting, weather, and interference from dynamic objects may all have an impact on the accuracy of the entire system during autonomous positioning and during the navigation of mobile visual simultaneous localization and mapping (SLAM) robots. As it is an essential element of visual SLAM systems, loop closure detection plays a vital role in eradicating front-end-induced accumulated errors and guaranteeing the map's general consistency. Presently, deep-learning-based loop closure detection techniques place more emphasis on enhancing the robustness of image descriptors while neglecting similarity calculations or the connections within the internal regions of the image. In response to this issue, this article proposes a loop closure detection method based on similarity differences between image blocks. Firstly, image descriptors are extracted using a lightweight convolutional neural network (CNN) model with effective loop closure detection. Subsequently, the image pairs with the greatest degree of similarity are evenly divided into blocks, and the level of similarity among the blocks is used to recalculate the degree of the overall similarity of the image pairs. The block similarity calculation module can effectively reduce the similarity of incorrect loop closure image pairs, which makes it easier to identify the correct loopback. Finally, the approach proposed in this article is compared with loop closure detection methods based on four distinct CNN models with a recall rate of 100% accuracy; said approach performs significantly superiorly. The application of the block similarity calculation module proposed in this article to the aforementioned four CNN models can increase the recall rate's accuracy to 100%; this proves that the proposed method can successfully improve the loop closure detection effect, and the similarity calculation module in the algorithm has a certain degree of universality.
视角、光照、天气的变化以及动态物体的干扰，在移动视觉同步定位与建图（SLAM）机器人自主定位和导航过程中，都可能对整个系统的精度产生影响。作为视觉SLAM系统的关键要素，回环检测在消除前端累积误差和保证地图整体一致性方面发挥着至关重要的作用。目前，基于深度学习的回环检测技术更侧重于增强图像描述符的鲁棒性，而忽视了相似性计算或图像内部区域之间的联系。针对这一问题，本文提出了一种基于图像块相似性差异的回环检测方法。首先，使用具有有效回环检测功能的轻量级卷积神经网络（CNN）模型提取图像描述符。随后，将相似度最高的图像对均匀划分为图像块，并利用图像块之间的相似度重新计算图像对的整体相似度。块相似度计算模块能够有效降低错误回环闭合图像对的相似度，从而更易于识别正确的回环。最后，将本文提出的方法与基于四种不同CNN模型的回环检测方法进行比较，召回率准确率达100%；该方法表现显著更优。将本文提出的块相似度计算模块应用于上述四种CNN模型，可将召回率准确率提高到100%；这证明了所提方法能够成功提高回环检测效果，且算法中的相似度计算模块具有一定的通用性。
