Human Activity Recognition Based on Convolutional Neural Network via Smart-phone Sensors
Zesheng Chen, Min Qi Zhou, Lichun Feng, Bingnan Li · 2022
To resolve the problem of insufficient accuracy in human activity recognition based on a single accelerometer sensor and traditional machine learning, this paper collects data on human activities using a smartphone embedded with multisensors, and then develop a framework based on a convolution neural network to classify human activities. When building a four- layer neural network, maximum pooling is utilized for every two layers, and the dropout technique is employed in case of overfitting. A full connection layer is then built using average pooling, and the Softmax approach is used for multi- classification. Experiments reveal that the suggested framework of convolution neural network enhances the accuracy of human activity recognition significantly.