Visual scene understanding and place recognition are the most challenging problems that mobile robots must solve for to achieve autonomous navigation. To reduce the high computational complexity of many global optimal search strategies, a new two-stage loop closure detection (LCD) strategy is developed in this paper. The front-end sequence node level matching (FSNLM) algorithm is based on the local continuity constraint of the motion process, which avoids the blind search for the global optimal match, and matches the image nodes via a sliding window to accurately find the local optimal matching candidate nodesets. In addition, the back-end image level matching (BILM) algorithm combined with an improved semantic model, DeepLab_AE, uses a convolutional neural network (CNN) as a feature detector to extract visual descriptors. It replaces traditional feature detectors that are manually designed by researchers in the computer vision field and cannot be applied to all environments. Finally, the performance of the two-stage LCD algorithm is evaluated on five public datasets, and is compared with the performance of other state-of-the-art algorithms. The evaluation results prove that the proposed method compares favorably with other algorithms.

视觉场景理解和位置识别是移动机器人实现自主导航所必须解决的最具挑战性的问题。为了减少许多全局最优搜索策略的高计算复杂度，本文开发了一种新的两阶段闭环检测（LCD）策略。前端序列节点级别匹配（FSNLM）算法基于运动过程的局部连续性约束，避免了对全局最优匹配的盲目搜索，并通过滑动窗口对图像节点进行匹配以准确地找到局部最优匹配候选节点集。此外，后端图像级别匹配（BILM）算法与改进的语义模型DeepLab_AE相结合，使用卷积神经网络（CNN）作为特征检测器来提取视觉描述符。它取代了传统特征检测器，后者是计算机视觉领域的研究人员手动设计的，无法应用于所有环境。最后，在五个公共数据集上评估了两阶段LCD算法的性能，并将其与其他最新算法的性能进行了比较。评估结果表明，该方法与其他算法相比具有优势。