Fig. 3: Full-body and limb-focused classification.

a Comprehensive diagram illustrates the process of motion data, which undergoes signal processing, feature extraction, and classification through the model's structure. b Loss minimization graph of the training and validation sets for full body motion classification, accompanied by a scatterplot illustrating the validation and training accuracy for their respective validation sets. c Confusion matrix illustrating the test set results of full body motion. d Precision–recall curves comparing four classification models on the full-body motion dataset. The Multimodal CNN model processes three separate CWT-based inputs from each axis using a multi-branch architecture. The RMS CNN model applies a root-mean-square operation across the x, y, and z-axis and generates a single scalogram image for CNN-based classification. In contrast, the SVM and Logistic Regression models directly use the raw time-series sensor data as input. e Classification accuracy of the Multimodal CNN and RMS CNN models at three sensor attachment positions: –25% (gray), 0% (black), and +25% (red) relative to the center of the limb segment. Positions are defined based on the distance from the wrist to the elbow (arm) and from the ankle to the knee (leg). f Time consumed for each trial, consisting of the time taken for the application to send the motion signal to the remote PC and the period for the computer to process and return the classification result back to the mobile device through the cloud server.