Human Activity Recognition Algorithm Based on ACNN-GRU

Abstract

In the field of human activity recognition using wearable sensors, the accurate extraction of features from raw data and the development of an appropriate classification model are crucial for improving recognition accuracy. To respond to this and other challenges, we are put forward an ACNN-GRU deep learning model that amalgamate a convolutional neural network, a gated recursive unit, and an attention mechanism. The Convolutional Neural Network (CNN) virtually catch the spatial information in the sensor data, while the Gated Recursive Unit (GRU) models the temporal correspondence in the data sequence. By incorporating the attention mechanism, our model automatically learns and focuses on the salient features essential to activity recognition, resulting in more efficient extraction of fine-grained features from sensor data. Our method well identified six everyday activities: Walking, Standing, Sitting, Laying, Upstairs and Downstairs. Experimental results on the UCI open-source dataset demonstrate an impressive recognition rate of 97.65%, surpassing the performance of traditional classification models.