SLAM Loop Closure Detection and Verification based on the Improved Siamese Network

Abstract

Visual Simultaneous Localization and Mapping (VSLAM) occupies a pivotal position in the robotics field. The loop closure detection module, which is related to the quality of mapping and the accuracy of positioning, is an indispensable part of SLAM. In recent years, neural networks are often used to replace the feature extraction part of loop closure detection. These methods can extract more helpful features, but the effect is not significant. In this paper, an improved siamese network is proposed to view the loop as a whole to improve the real-time performance of SLAM. Firstly, an improved 2D-siamese is proposed to obtain candidate key frames. In order to integrate feature extraction and similarity comparison, this 2D-siamese uses SE-Resnet network as its branch. Secondly, a 3D-siamese network, which verifies the continuity by using continuous images, is proposed for eliminate mismatches and improve loop closure detection accuracy. The experimental results on TUM and KITTI datasets show that the proposed method can greatly improve the accuracy and recall rate of the loop closure detection.