EXPLAINABLE IMU-BASED SURFACE CLASSIFICATION FOR HUMAN-CENTERED DECISION SUPPORT IN MOBILE ROBOT OPERATIONS

Abstract

The surface recognition can assist in the navigation and management of a mobile robot, but merely a classification accuracy cannot fully justify why the particular terrain classification was made. This paper analyses a leakage aware and explainable classical machine learning pipeline based on the Surface Type Classification for Autonomous Robot Indoor Navigation dataset. In total, 7,626 time-series segments from nine different indoor surfaces are used. Four channels with orientation quaternion components are omitted due to their possible dependence on trajectory. Six channels, namely, the angular velocity and linear acceleration components are selected. 316 statistical, temporal, autocorrelation, magnitude, and frequency domain features are computed. Nine classical models are considered applying 3-fold StratifiedGroupKFold cross-validation in order to preserve all examples of the same recording group in one fold. Since the class distribution is highly imbalanced, macro F1 score is considered to be the main performance measure. The engineered feature-based RBF SVM performed best, achieving 0.6123 Macro F1, 0.6216 Macro Recall, 0.6446 Accuracy, and 0.9276 Macro ROC-AUC. Global explanation is generated using permutation importance in held-out cross-validation, whereas Kernel SHAP is applied to held-out examples for feature contribution inspection. The contribution is the combination of surface classification and explanation pipelines, which allows converting robot sensors predictions into human-readable information for further engineering and human-centred operations.