Reliable Loop Closure Detection Using 2-channel Convolutional Neural Networks for Visual SLAM

To improve the reliability of identifying revisited places in cross-session visual SLAM, a reliable loop closure detection method is proposed in this paper. The proposed approach is based on 2-channel convolutional neural network for estimating image-to-image similarity. Transfer learning is employed, which enables a small dataset of specific scenes to be used to fine-tune the pre-trained 2-channel convolutional neural network. The posterior probability of continuous loop closure detection is estimated using Bayes filters with memory management to improve the real-time performance. Experiments on the cross-season and cross-time datasets are also demonstrated to evaluate the performance of the proposed method.