Walking-posture Classification from Single-acceleration-sensor Data using Deep Learning

Abstract

We described a walking-posture classification method from a single accelerator attached to a human waist using a deep learning technique. We considered deep learning architectures for a single accelerator based on previous human activity recognition studies and investigated the classification accuracy of the proposed method using the walking-posture dataset. The results demonstrate that a deep learning approach to walking-posture classification using a single accelerator is more useful than the conventional SVM approach. Additionally, we also confirmed that a hybrid network architecture with three convolutional neural layers, two pooling layers between the convolutional layers, and a long short-term memory layer achieved the best accuracy of 0.9803 compared to other network architectures. We also confirmed the deep learning approach yielded more correct classification for each walking-posture category in spite of the difficulty to detect the classification by the SVM approach.