MATC-Net: Learning compact sequence representation for hierarchical loop closure detection

Abstract

Loop closure detection (LCD) is a challenging task to judge whether the current position of an intelligent robot returns to the previously visited position. Mainstream appearance-based approaches apply robust image representation techniques to describe the scene. However, most of these methods are designed for single images, and the sequence representation method incorporating temporal sequence information is still in the preliminary exploration. In this paper, we propose a compact sequence representation method for hierarchical LCD, ensuring conspicuous performance for the LCD task. Deriving from a group of global features of image sequences, we propose a multi-scale asymmetric temporal convolution network (MATC-Net), which generates sequential features and transformed global features through its aggregation branch and transformation branch, respectively. Based on these two types of features, a MATC-Net-based hierarchical LCD framework including two similarity measurement processes is constructed, through which the best place match is identified. The experimental results show that our method outperforms other counterparts on three datasets, exhibiting the promising potential of leveraging sequential features to LCD task.