A sequence-based visual place recognition method for aerial mobile robots

Abstract

Recent developments in sequence-based visual place recognition (VPR) have shown robust localization capability on ground vehicles; however, deploying existing sequence-based VPR techniques on aerial mobile robots is still challenging, as topologically ordered database is not available for aerial agents. In this paper, we develop a new sequence-based VPR algorithm, which is applicable to aerial mobile robots. The proposed method models the agent’s position as a probability function conditioned on all the visual observations and accumulates the position belief as the observation increases, which helps to filter out the inconsistent false recognitions. To testify the effectiveness of the proposed method, we have conducted series of experiments under different visual environments. The experimental results show that, by using a remote sensing image as reference and onboard downward-looking camera images as observations, the proposed method can robustly estimate the position of the agent and have a localization rate 30% higher than the single-image-based VPR algorithm.