关键词: 特征提取, 卷积神经网络, 双向门控循环单元, 人类活动识别, 深度学习

Abstract:

Deep learning based HAR （Human Activity Recognition） methods suffer from the complexity of manual feature extraction process and the difficulty of mining complex time-series dependencies when dealing with time-series data. How to effectively and automatically extract multi-scale features of human activities and mine the correlation features of time series is the key factor to improve the accuracy of HAR. To address the above issues， a Multi-scale one-Dimensional Convolutional Neural Network Bidirectional Gated Recurrent Unit （1DMCNN BiGRU） model was proposed. Refined perceptual signal features were extracted by multi-scale convolution， while correlation features of the overall signal before and after extracted by BiGRU （Bidirectional Gated Recurrent Unit） were fused to improve the recognition accuracy of the model. Experimental results on real-life scene datasets USC-HAD， WISDM， and PAMAP2 show that compared to the suboptimal CNN-LSTM （Convolutional Neural Network with Long Short-Term Memory） model， the proposed model has improved the recognition accuracy by 1.06%， 1.23% and 1.71%， respectively， with high recognition accuracy， which verifies the effectiveness of the proposed model for HAR.

Key words: feature extraction, Convolutional Neural Network (CNN), Bidirectional Gated Recurrent Unit (BiGRU), Human Activity Recognition (HAR), deep learning