Performance Evaluation of Deep Neural Networks in Detecting Loop Closure of Visual SLAM

This paper concerns the problem of Loop Closure Detection (LCD) of visual Simultaneous Localization and Mapping (SLAM). The LCD is a crucial model to reduce the accumulative error in visual SLAM. The traditional LCD methods use hand-crafted features, which ignore useful information. We propose a LCD method based on Convolutional Neural Networks (CNNs) without any manual intervention for visual features. We compare and analyze several popular deep neural networks models for LCD. Two open datasets has been used to evaluate the performance of LCD in terms of mean-per-class accuracy. The results show that deep neural networks are feasible for LCD and the ResNet50 network outperforms the other deep neural networks.