While robotic mapping of static environments has been widely studied, life-long mapping in non-stationary environments is still an open problem. We present an approach for long-term representation of natural environments, where many of the observed changes are caused by pseudo-periodic factors, such as seasonal variations, or humans performing their daily chores.
Rather than using a fixed probability value, our method models the uncertainty of the elementary environment states by their frequency spectra. This allows to integrate sparse and irregular observations obtained during long-term deployments of mobile robots into memory-efficient models that reflect the recurring patterns of activity in the environment. The frequency-enhanced spatio-temporal models allow to predict the future environment states, which improves the efficiency of mobile robot operation in changing environments. In a series of experiments performed over periods of weeks to years, we demonstrate that the proposed approach improves mobile robot localization, path and task planning, activity recognition and allows for life-long spatio-temporal exploration.