WiMTI: A multitask learning model for WiFi-based identity and posture recognition.

Abstract

WiFi-based identity recognition has attracted considerable attention due to its nonintrusive nature and ability to preserve privacy, unlike traditional camera-based or wearable-sensor-based methods. However, variations in human posture can severely disrupt the stability of WiFi signal characteristics, notably reducing the accuracy and reliability of WiFi-based identity recognition systems. Therefore, we propose WiMTI, a novel multitask learning (MTL) model designed for simultaneous identity and posture recognition. By explicitly capturing posture-related features through MTL, WiMTI effectively reduces the negative impact of posture variations on identity recognition. Specifically, WiMTI employs dynamic cross-stitch units for adaptive feature fusion and integrates fractal dimension analysis to enhance feature representation. Experimental results show that WiMTI achieves state-of-the-art average accuracies of 98.25% for identity recognition and 92.83% for posture recognition. These results demonstrate WiMTI's robustness and effectiveness, making it suitable for practical applications such as access control and caregiving.