Implementation of Hybrid Deep Learning Architecture on Loop-Closure Detection

Abstract

The loop-closure detection is an essential part in visual SLAM. This paper proposes an effective approach to generate high-level semantic image features for improving loop-closure detection and place recognition, based on a hybrid Convolutional Neural Network (CNN) modified by us to especially cope with real-time feature extraction. This hybrid deep learning architecture is adapted from ResNet-18 and enhanced by split-transform-merge idea as well as squeeze-and-excitation structure, such that the network’s ability of image representation can be compensated without compromising on efficiency. Then, we design a simple yet well-fitted compression strategy for reducing the dimension of deep convolutional features. The strategy significantly accelerates the speed of image-pairs matching with little loss or even improvements in Precision/Recall performance. Our approach is validated on public benchmark datasets and compared to state-of-the-art methods. Experiments demonstrate improved Precision/Recall outperforming state of the art. Additionally, that precision declines very slow with the increasing of recall indicates potential of further progress.