Spectral saliency model for an appearance only SLAM in an indoor environment

摘要

Simultaneous Localization and Mapping, SLAM, is a prime requirement for autonomous navigation and it is an integral part of a modern robot. An incremental appearance only SLAM based on non-quantized local features drastically suffers from large number of landmarks detected in the captured imagery. In order to decrease the number of landmarks, we propose a visual saliency model that is used as first stage in a SLAM algorithm; whereby it detects a small region (25% of image) in the captured imagery. This small region consists of those salient landmarks which are most conspicuous in the surrounding and brain gives more important to them. Only landmarks lying within the selected small salient region of the image are considered for use in the SLAM algorithm. The saliency stage of the algorithm we present is based on a spectral visual saliency model. We analyze its efficacy in terms of Receiver Operating Characteristic, ROC, and shuffled Area Under the Curve, sAUC, based on human fixation data tested on a popular dataset. In the second stage a tailored form of an existing SLAM framework is used to verify the applicability of the proposed saliency stage in a SLAM application. Finally, we use an indoor dataset and with numerous simulations of different parameters of SLAM to show the effectiveness of the proposed approach.