Learning Condition-invariant Scene Representations for Across the Seasons Place Recognition

H. Biswas

For autonomous navigation in an unknown environment, a mobile robot has to build a map of the environments and localize itself within the map. When a robot movies in an environment, the errors in motion accumulate over time, which gradually makes the map inconsistent. To reduce the error in estimating the map of the environment, it is necessary for a robot to recognize previously visited locations, called as a loop-closure detection. When an environment goes through extreme perceptual changes, e.g., seasonal changes, it becomes a harder problem for a mobile robot to recognize its previously visited locations. The purpose of this research work is to develop an algorithm that learns the robust representation of the scenes in presence of the seasonal changes to assist the correct loop-closure detection. To reach this goal, independent component analysis and autoencoder have been adopted. In this work, it has been observed that, the ICA has the promising capabilities of extracting the condition-invariant image descriptors when a robot is operating across the wide range of appearance changing environment. The potential of distinguishing the condition-variant and condition-invariant features of a scene by an autoencoder has been investigated, which shows a considerable performance and discovers a possible route to take this work to the next step. The performance is evaluated with the baseline method using the precision-recall curves and fraction of correct matches. For the selected evaluation criteria, our algorithm showed superior performance compared to the baseline method.