A Lightweight Neural Network for Loop Closure Detection in Indoor Visual SLAM

Abstract

Loop closure detection (LCD) plays an important role in visual simultaneous location and mapping (SLAM), as it can effectively reduce the cumulative errors of the SLAM system after a long period of movement. Convolutional neural networks (CNNs) have a significant advantage in image similarity comparison, and researchers have achieved good results by incorporating CNNs into LCD. The LCD based on CNN is more robust than traditional methods. As the deep neural network frameworks from AlexNet and VGG to ResNet have become smaller while maintaining good accuracy, indoor LCD does not need robots to finish a large number of complex processing operations. To reduce the complexity of deep neural networks, this paper presents a new lightweight neural network based on MobileNet V2. We propose a strategy to use Efficient Channel Attention (ECA) to insert into Compressed MobileNet V2 (ECMobileNet) for reducing operands while maintaining precision. A corresponding loop detection method is designed based on the average distribution of ECMobileNet feature vectors combined with Euclidean distance matching. We used TUM datasets to evaluate the results, and the experimental results show that this method outperforms the state-of-the-art methods. Although the model was trained only on the indoorCVPR dataset, it also demonstrated superior performance on the TUM datasets. In particular, the proposed approach is more lightweight and highly efficient than the current existing neural network approaches. Finally, we used TUM datasets to test LCD based on ECMobileNet in PTAM, and the experimental results show that this lightweight neural network is feasible.