Triplet loss based metric learning for closed loop detection in VSLAM system

Abstract

Closed loop detection can alleviate the error accumulation during the operation of Visual Simultaneous Localization and Mapping (VSLAM) system, which is of great significance to the accuracy and robustness of the robot. Triplet loss based metric learning has been proposed for closed loop detection in this paper. Firstly, a triplet selection strategy has been constructed. The Softplus function is applied to triplet loss so that the loss in the negative axis will be soft margin. The adaptive margin has been proposed to maintain the mapping distribution in metric space. Metric learning converts keyframes into feature vectors, evaluating the similarity of keyframes by calculating the Euclidean distance between feature vectors, which is utilized to determine whether a closed loop is formed. Secondly, detection strategy of candidate keyframes for loop detection is introduced according to Euclidean distance. Finally, triplet loss based metric learning is applied to closed loop detection for VSLAM system. VSLAM datasets have been applied to evaluate the precision and recall of metric learning model, and then the established VSLAM system has been implemented in real-time application. The experimental results illustrate the feasibility and effectiveness of proposed method, which can be further applied to practical VSLAM systems.