Assessing Impact of Sensors and Feature Selection in Smart-Insole-Based Human Activity Recognition

Abstract

Human Activity Recognition (HAR) is increasingly used in a variety of applications, including health care, fitness tracking, and rehabilitation. To reduce the impact on the user's daily activities, wearable technologies have been advanced throughout the years. In this study, an improved smart insole-based HAR system is proposed. The impact of data segmentation, sensors used, and feature selection on HAR was fully investigated. The Support Vector Machine (SVM), a supervised learning algorithm, has been used to recognise six ambulation activities: downstairs, sit to stand, sitting, standing, upstairs, and walking. Considering the impact that data segmentation can have on the classification, the sliding window size was optimised, identifying the length of 10 s with 50% of overlap as the best performing. The inertial sensors and pressure sensors embedded into the smart insoles have been assessed to determine the importance that each one has in the classification. A feature selection technique has been applied to reduce the number of features from 272 to 227 to improve the robustness of the proposed system and to investigate the importance of features in the dataset. According to the findings, the inertial sensors are reliable for the recognition of dynamic activities, while pressure sensors are reliable for stationary activities; however, the highest accuracy (94.66%) was achieved by combining both types of sensors.

Keywords: activity recognition; feature selection; machine learning; smart insole; window size optimisation.

Figure 1

ActiSense Kit, IEE Luxembourg S.A. (a) IEE Smart Foot Sensor (b) Example of how the system is worn by the user.

Figure 2

Result of the grid search strategy to identify the optimal sliding window size for classification. The 10 s window size with 50% of overlap achieved the highest F1-Score (94.64%). Overall in almost all the window sizes tested, the introduction of the overlapping between consecutive windows allowed an increase in the classification performance.

Figure 3

Evaluation of ROC curves for each activity using three different datasets built from different types of sensors: Inertial sensors and pressure sensors (I + P), pressure sensors (P), and inertial sensors (I). In (e), since the sitting activity is correctly classified in all the experiments, the curve “I+P: AUC (1.00)” overlaps “I: AUC (1.00)” and is not visible.

Figure 4

Number of features selected by each foot for each sensor. The hallux is the most relevant pressure sensor (PS), whereas, the accelerometer (Acc) is the most significant inertia sensor, followed by the gyroscope (Gyr). The x-axis of the magnetometer (Mag) is the only sensor in this investigation that had no bearing on the activity classification.