Loop closure detection for visual SLAM using PCANet features

Loop closure detection benefits simultaneous localization and mapping (SLAM) in building a consistent map of the environment by reducing the accumulate error. Handcrafted features have been successfully used in traditional approaches, whereas in this paper, we show that unsupervised features extracted by deep learning models, can improves the accuracy of loop closure detection. In particular, we employ a cascaded deep network, namely the PCANet, to extract features as image descriptors. We tested the performance of our proposed method on open datasets to compare with traditional approaches. We found that the PCANet features outperform state-of-the-art handcrafted competitors, and are computational efficient to be implemented in practical robotics.