SMCN: Simplified mini-column network for visual place recognition

Abstract

Visual place recognition(VPR) is challenging because the places to be recognized are often affected by complex environmental changes. It is important for loop closure detection in SLAM. In recent years, a large range of approaches have been developed to address this challenge. Among these approaches, the use of biological heuristics and structural information have improved the VPR performance. In this paper, we attempt to improve the biological heuristic method MCN, and combine the available but often overlooked structural information intra-set similarity to propose our first algorithm SMCN. Furthermore, we propose a novel temporal filter that considers temporal continuity and combines it with SMCN to get our second algorithm SMCNTF. Evaluation of both algorithms on ten dataset combinations shows that, our best model SMCNTF has a maximum increase of 32% in average precision and at least a 100-fold increase in computing efficiency. Moreover, fewer parameters need to be tuned comparing with MCN. Generally, our algorithm is much better than MCN.