Towards improving the efficiency of sequence-based SLAM

We propose a method in this paper to perform sequence-based appearance SLAM in an efficient and effective way. Sequence-based SLAM (or SeqSLAM for short) makes use of the image descriptors extracted from a series of consecutive frames and matching is done between two such image sequences. It has been shown to be effective in dealing with significant illumination change where localization and mapping can be conducted under different time periods and weather conditions. To address the computational issue that can arise from the exhaustive search of the candidate sequences with the increase of map size, we use a particle filter to implement the Bayes filtering framework of estimating the true match. The resampling of the particles allows us to maintain only a small number of hypotheses while still capturing the true distribution of the robot location. Our method is highly scalable and efficient, validated on a large dataset with comparable results to the original algorithm in terms of performance.