A Comparison of Wearable Sensor Configuration Methods for Human Activity Recognition Using CNN

Abstract

The number and location configuration methods of wearable sensors for human activity recognition (HAR) are analytically discussed. Based on the publicly available Daily and Sports Activities data set, a convolutional neural network (CNN) was built to recognize 19 kinds of daily and sports activities, and then the model was optimized for better performance. The results of numerous comparative experiments show that deep learning-based HAR is better than machine learning-based HAR in terms of accuracy, and its improvement in accuracy is not directly related to the increase of sensor quantity. Due to its strong capability of feature extraction, deep learning extracts not only activity-related features but also individual differences, therefore, the location with less individual randomness should be selected according to practical engineering. Moreover, the results are also influenced by the limb symmetry in the data set. Finally, the feasibility of achieving higher accuracy with fewer sensors is proved.