2 July 2021 in IEEE Robotics and Automation Letters
Loop closure detection (LCD) is an important portion of Simultaneous Localization and Mapping (SLAM) because of its ability to reduce accumulated position errors. In this letter, we propose a novel loop closure detection algorithm named ESA-VLAD. The crucial part of ESA-VLAD is a redesigned network with EfficientNetB0 as backbone for extracting global features, which integrates a second-order attention module in order to effectively learn the correlations between features within the feature map. A trainable Vector of Local Aggregated Descriptors (NetVLAD) is integrated in the last layer of the network to generate a compact and fixed-length global feature. Knowledge distillation strategy is adopted in training of the proposed network to accelerate the training process. For the global features, Hierarchical Navigable Small World (HNSW) is employed to retrieve the loop closure candidate images. In addition, an efficient geometrical consistency check based on local difference binary (LDB) descriptors is designed to verify loop closure matches. Experiments on several public datasets demonstrate that ESA-VLAD can obtain higher recall rates under 100% precision and less processing time per frame compared to other typical and state-of-the-art methods.
31 December 2020 in Electronics Letters
Loop‐closure detection is important in large‐scale localisation system. However, it is still difficult in a dynamic environment. An online fast loop‐closure detection algorithm based on Deeplabv3 with MobileNetV2 (DpMn2) as network backbone and local difference binary descriptor is proposed, and the algorithm is named as DpMn2‐LDB. DpMn2 splits out common dynamic objects of images, and then uses visual geometry group network (VGG16) that is trained on place‐centric data to extract global features for nearest neighbour image retrieval. The loop‐closure matches are verified based on LDB descriptors and random sample consensus (RANSAC). Experimental results show that the proposed method can obtain a higher recall rate under 100% precision with less execution time per frame on several public datasets compared with other typical or state‐of‐the‐art algorithms.