Human Activity Recognition Using Machine Learning in Wireless Body-Area Networks

This work evaluates whether the Bluetooth Low Energy (BLE) Received Signal Strength Indicator (RSSI) alone can support human-action recognition (HAR) in a star-topology Wireless Body-Area Network (WBAN). Nine participants performed three activities (standing, walking, and sitting) while wearing a chest-mounted central node and four peripheral nodes on the wrists and ankles. RSSI streams were segmented into 10 s windows; from each window, fourteen statistical descriptors (e.g., mean, variance, percentiles) were extracted. Six conventional classifiers (k-NN, SVM, Random Forest, Logistic Regression, Naïve Bayes, and Gradient Boosting) were evaluated using 30 random seeds. The Gradient-Boosting model achieved the highest accuracy, 93% ± 1.46. These demonstrate that RSSI-based features from only two sensors provide sufficient information for low-complexity accurate HAR in a WBAN.