Cross-Domain Recognition Algorithm for Fitness Activities and Action Frequency Based on Dual Attention Prototype Networks

With the acceleration of life pace and improvement of living standards, adhering to exercise and fitness has become a common way for modern individuals to pursue a healthy lifestyle. Some fitness activity recognition algorithms based on channel state information (CSI) have been proposed to monitor the fitness effect. However, action frequency is also crucial for fitness effectiveness. The existing algorithms only recognize fitness activities without recognizing action frequency, and can only achieve satisfactory recognition accuracy in the trained experimental scenarios. To address these issues, we propose a cross-domain recognition algorithm for fitness activities and action frequency based on dual attention prototype networks, abbreviated as CDR-DAPNet algorithm. This algorithm uses CSI-ratio method to eliminate the random offset of CSI phase information, and converts the phase information into the form of an image, which is input into the constructed cross-domain recognition network. In the proposed recognition network, we construct a feature extraction module with serial structure consisting of spatial attention mechanism, ResNet34 network, and channel attention mechanism, to extract and reconstruct features from CSI phase images. Then, a recognition module with parallel structure consisting of two prototype networks is built for recognizing fitness activities and action frequency. Experimental results demonstrate that the CDR-DAPNet algorithm achieves high average recognition accuracy within the domain, and has good cross-domain performance when the amount of training data is sufficient or the training is assisted with few samples from cross-domain scenes.