심층 신경망을 활용한 심전도 신호 기반 정적 행동 인식

초록

This study focuses on the classification of human static activities using a single wearable electrocardiogram(ECG) sensor. Four static activities(lying, sitting, standing, and squatting) were defined for the classification task. A wearable ECG sensor was attached to the V2 position of participants for data collection. The ten participants performed the defined static activities in a random order during the experiment. Deep neural networkbased models, specifically a onedimensional convolutional neural network(1D CNN) and hybrid models combining 1D CNN with a long-short term memory(LSTM) or gated recurrent unit(GRU) layers, were utilized for classification. Comparing the classification performance of the models, the 1D CNN-GRU hybrid model achieved the highest test accuracy(f1-score) of 91.45%(91.48%), representing a improvement over the 1D CNN model and 1D CNNLSTM hybrid model. Additionally, the t-SNE results demonstrated a more clustered distribution with the hybrid models. In conclusion, improved classification accuracy and more distinct data distribution patterns indicate the effectiveness of the hybrid model in accurately identifying static activities. Therefore, the proposed hybrid model in this study enhances the potential for nonintrusive and user-friendly activity recognition.

키워드