TLCD: A Transformer based Loop Closure Detection for Robotic Visual SLAM

Chenghao Li, Hongwei Ren, Minjie Bi, Chenchen Ding, Wenjie Li, Rumin Zhang, Xiaoguang Liu, Hao Yu

Abstract: Loop closure detection (LCD) can effectively correct errors in visual odometry. It is thereby a critical part in robotic visual simultaneous localization and mapping (SLAM) system, which is widely used in modern robotic systems such as sweeping robots and drones. In this paper, we propose a transformer-based loop closure detection algorithm (TLCD), which employs a distillation transformer as backbone to extract global features, and is combined with a sequence matching as back-end processing of principal component analysis (PCA) algorithm. TLCD can accurately provide Precision-Recall curve based on several public datasets including CityCentre and New-College datasets. Results show that TLCD’s average accuracy is up to 16.91% higher than the traditional LCD method. It is also about 3.18% higher accuracy than the state-of-the-art convolutional neural network (CNN) based LCD method.