Are We There Yet? Challenging SeqSLAM on a 3000 km Journey Across All Four Seasons

When operating over extended periods of time, an autonomous system will inevitably be faced with severe changes in the appearance of its environment. Coping with such changes is more and more in the focus of current robotics research. In this paper, we foster the development of robust place recognition algorithms in changing environments by describing a new dataset that was recorded during a 728 km long journey in spring, summer, fall, and winter. Approximately 40 hours of full-HD video cover extreme seasonal changes over almost 3000 km in both natural and man-made environ- ments. Furthermore, accurate ground truth information are provided. To our knowledge, this is by far the largest SLAM dataset available at the moment. In addition, we introduce an open source Matlab implementation of the recently published SeqSLAM algorithm and make it available to the community. We benchmark SeqSLAM using the novel dataset and analyse the influence of important parameters and algorithmic steps.