Joint Behavior and Location Recognition Framework Based on Electromagnetic Fingerprints

Simultaneous perception of human behavior and location in indoor environments is a significant challenge. Multi-task Learning (MTL) frameworks have been shown to effectively address this problem by leveraging correlations between tasks. However, traditional MTL models usually rely on a fixed parameter sharing mechanism, which can limit model learning capabilities and lead to substantial accuracy variations across tasks. To address these issues, we propose a novel joint perception framework that utilizes Channel State Information (CSI) finger-printing. First, we introduce a selective sharing method based on sparse parameters to mitigate the problems associated with fixed parameter sharing in MTL. This approach dynamically allocates shared parameters according to the specific needs of each task, thereby enhancing the flexibility of the model. Second, to further balance the model performance between two tasks, we introduce an adaptive loss weight adjustment approach. This approach dynamically adjusts the loss weights based on the performance of each task, ensuring good accuracy for both behavior recognition and location estimation. Experimental results demonstrate that our proposed framework significantly enhances accuracy in both behavior recognition and location estimation.