A Framework for Automatic Selection of Indoor Landmarks using Machine Learning Algorithms and Shapley Additive Explanations

Abstract

Landmarks are salient objects in an environment compared to their surroundings. However, a challenge of landmark-based navigation is selecting the most salient landmarks to include in route instructions. Current approaches mainly adopt weighted linear models, which assume that landmarks have absolute salience values. However, this contradicts the definition of landmarks as being salient in comparison to their surroundings. In this work-in-progress study, a probability-based soft classification approach is proposed to automatically select indoor landmarks. Specifically, we aggregated fundamental salience measures regarding visual, structural, and semantic dimensions from related studies to create an indoor landmark dataset. Then we, compared the performances of machine learning classifiers with several metrics and interpreted the local contributions of salience measures. Finally, we utilized a probability calibration technique that allows for finer-grained representations of indoor landmarks to include them in the route guidance process. According to the preliminary results of this study, boosting-based machine learning algorithms provide remarkable results, and functional uniqueness, category, and intensity measures are considered more important to select indoor landmarks. Moreover, our soft probability-based classification framework seems promising for selecting and representing landmarks in a fine-grained manner. However, the feasibility of the proposed framework should be further validated with user studies.