Loop Closure Detection in Visual Appearance-Based SLAM Using Deep Autoencoders

Loop closure detection (LCD) and trajectory generation are critical components of visual simultaneous localization and mapping (vSLAM). In this paper, we aim to solve the LCD and trajectory generation problem in vSLAM using a newly devised vector quantization (VQ) algorithm. The proposed new VQ algorithm is constructed based on a selfsupervised deep convolutional autoencoder (AE). The new VQ step is then incorporated into the two famous SLAM algorithms fast appearance-based mapping (FABMAP) and ORB-SLAM, which we now call AE-FABMAP and AE-ORB-SLAM, respectively. Experiments show that using self-supervised autoencoders in the VQ step is far more efficient in terms of speed and memory consumption with respect to other methods such as graph convolutional neural networks. Furthermore, the newly presented algorithms, AE-ORB-SLAM and AE-FABMAP outperform the standard FABMAP2 and ORB SLAM, and in large-scale SLAM, the new approaches improve the accuracy and recall of the LCD.