Loop closure detection (LCD) is crucial for the simultaneous localization and mapping system of an autonomous robot. Image features from a convolution neural network (CNN) have been widely used for LCD in recent years. Instead of directly using the feature vectors to compute the image similarity, we propose a novel and easy-to-implement method that manages features from a CNN via a novel approach to improve the performance. In this method, the elements of feature maps from the higher layer of the CNN are clustered to generate CNN words (CNNW). To encode spatial information of CNNW, we create word pairs (CNNWP) that are based on single words to improve the performance. In addition, traditional tricks that are used in methods that are based on bag of words (BoW) are integrated into our approach. We also demonstrate that the feature maps from lower layers can be used as descriptors to conduct local region matching between images. Via this approach, we can perform geometric verification for possible loop closures, similar to BoW methods, in our approach. The experimental results demonstrate that our method substantially outperforms state-of-the-art methods that directly use CNN features for LCD.

中文翻译：

---

使用CNN单词进行闭环检测

闭环检测（LCD）对于自主机器人的同时定位和制图系统至关重要。近年来，来自卷积神经网络（CNN）的图像特征已广泛用于LCD。代替直接使用特征向量来计算图像相似度，我们提出了一种新颖且易于实现的方法，该方法通过一种新颖的方法来管理CNN的特征，从而提高了性能。在这种方法中，来自CNN较高层的特征图元素被聚类以生成CNN词（CNNW）。为了对CNNW的空间信息进行编码，我们创建了基于单个单词的单词对（CNNWP）以提高性能。此外，在基于词袋（BoW）的方法中使用的传统技巧已集成到我们的方法中。我们还证明了来自较低层的特征图可以用作描述符来进行图像之间的局部区域匹配。通过这种方法，我们可以像BoW方法一样对可能的闭环进行几何验证。实验结果表明，我们的方法大大优于直接将CNN功能用于LCD的最新方法。