A window-level benchmark study on triaxial smartphone sensor fusion for human activity recognition using classical machine learning

Abstract Human Activity Recognition using smartphone sensors supports applications in health monitoring, context-aware computing, and activity assessment. This study evaluates a fused accelerometer, magnetometer, and gyroscope representation for smartphone-based Human Activity Recognition. A balanced processed benchmark containing 10,000 instances was represented using 45 handcrafted statistical and inter-peak-distance features extracted from the three sensor-magnitude signals. Eight classical machine-learning classifiers were evaluated in Orange Data Mining using stratified five-fold cross-validation at the window level. Classification Accuracy ranged from 98.66% for k-Nearest Neighbours to 100.00% for AdaBoost under the adopted benchmark. Fold-level uncertainty estimates and inferential model-comparison tests were unavailable; therefore, no statistically significant superiority among classifiers is claimed. The confusion matrices contain aggregated out-of-fold predictions generated during cross-validation. An audit of the processed matrix identified 1,046 unique feature-label rows among the 10,000 rows, indicating extensive exact repetition that may further increase optimism in row-level cross-validation estimates. Exploratory Random Forest importance scores and feature-distribution analyses show that several gyroscope-derived descriptors are highly ranked within the fused feature space. However, because separate sensor-subset models were not evaluated, the incremental performance attributable specifically to gyroscope inclusion cannot be quantified. The reported results characterise feature separability within the processed dataset and evaluation protocol. They do not establish generalisation to unseen participants or independent recording sessions because participant and session identifiers were unavailable in the processed feature matrix. Hardware latency, memory use, and energy consumption were not measured. The framework should therefore be considered a lightweight benchmark and a candidate for future deployment evaluation rather than a validated real-time mobile implementation.