26 August 2024 in IEEE Transactions on Instrumentation and Measurement
Non-periodic body movements is one of the most challenging issues in non-contact vital sign detection. The considerable and irregular displacements of the human body could corrupt the vital sign signals, drastically reducing detection accuracy. In this paper, an adaptive motion noise cancellation scheme based on Frequency-Modulated Continuous Wave (FMCW) virtual antenna array radar is proposed for real-time monitoring of vital signs with large body movements. This scheme suppresses the nonlinear noise caused by body motion and extracts accurate vital sign data. The proposed method is validated through simulation using a model of the vital sign detection system. Experiments were then carried out using a 77 GHz non-contact vital sign detection system, and the results were compared with data from a wearable device (MI6). The static experimental outcomes show that the errors in respiratory rate (RR) and heart rate (HR) are within 2 bpm and 3 bpm, respectively. The results of exercise test, with a velocity range of 0 to 0.5 m/s, show that the error of RR and HR is kept within 5bpm. The experimental results demonstrate the high efficiency of the algorithm in suppressing motion noise and accurately extracting RR and HR.
24 April 2025 in IEEE Sensors Journal
The classification and identification of human activities have been increasingly focused on in the fields of human-computer interaction, search and rescue, health detection, etc. Deep learning methods have been widely employed in target classification recognition. To enhance the classification performance, an attention-mechanism-based two-channel network (AB-TCN) is proposed. In this architecture, the attention module is embedded into the convolutional neural network to achieve feature enhancement and redundancy suppression in the spatial and channel domains. Furthermore, the short-window time-frequency image and long-window time-frequency image are separately input into two symmetrical channels for feature extraction and fusion to enhance the differential feature weight of target behavior. The method is simple and easy to implement, with low computational complexity. The experimental results show that the proposed method has higher detection accuracy, and the classification accuracy is increased by more than 5% compared with the traditional neural network architecture.
1 July 2026 in IEEE Sensors Letters