Encoding Accelerometer Signals as Images for Activity Recognition Using Residual Neural Network
share
Human activity recognition using a single 3-axis accelerometer plays an important and fundamental role in daily monitoring using wearable sensors and devices. In this paper, we address the recognition problem by encoding 3-channel accelerometer signals as images and using transferring learning approach. Though projecting time series onto image space is not a new topic, our method is the first time to encode 3-axis signals as 3-channels of color images using recurrence plots (RP) and train an 18-layer Residual Neural Network (ResNet-18) to do image classification, to the best of our knowledge. We also first propose a modified RP to overcome its tendency confusion problem, which has improved our system performance significantly. A user-independent database of over 1080 traces with 5 actions from 11 subjects is used to test our method. In addition, we evaluate its generalization ability on a public dataset. Experimental results have shown that our recognition framework achieves highly competitive accuracies with other state-of-the-art methods on both datasets. And it has a great potential to be applied for 3-axis accelerometer signals recognition.
READ FULL TEXT
