Skip to main content
2024 47th International Conference on Telecommunications and Signal Processing (TSP)

Full Program »
Video (.mp4)
View File
mp4
10.6MB

A Hybrid Residual CNN with Channel Attention Mechanism for Continuous User Identification Using Wearable Motion Sensors

Recognizing individuals continuously using wearable devices has several applications in personalized services and security. This paper presents a hybrid deep network architecture called SE-ResBiLSTM that integrates a squeeze-and-excitation (SE) mechanism with residual bidirectional LSTM layers. The goal is to enhance the accuracy of person identification using data from wearable motion sensors. The SE blocks facilitate channel-wise feature refinement, highlighting helpful information. Residual connections enable the training of deeper LSTM architectures. The proposed method's accuracy in identifying individuals across increasing periods between training and testing is evaluated using the SP-SW HAR dataset, which contains activity data collected from wristwatch sensors. By leveraging representation learning and exhibiting resilience against sensor noise and temporal fluctuations, SE-ResBiLSTM improves recognition accuracy compared to regular and convolutional LSTM networks. According to the trial data, SE-ResBiLSTM achieved a maximum identification accuracy of 98.71%. The experimental results demonstrate that this approach significantly boosts the performance of user identification using wristwatch sensors.

Sakorn Mekruksavanich
School of Information and Communication Technology, University of Phayao
Thailand

Wikanda Phaphan
King Mongkut’s University of Technology North Bangkok
Thailand

Anuchit Jitpattanakul
King Mongkut's University of Technology North Bangkok
Thailand

 

Privacy Policy

Powered by OpenConf®
Copyright ©2002-2024 Zakon Group LLC