Enhancing PIR-Based Multi-Person Localization Through Combining Deep Learning With Domain Knowledge

Abstract

Device-free localization (DFL) techniques based on pyroelectric infrared (PIR) sensors have attracted much attention due to their advantages of low cost, low power consumption, and privacy protection. However, existing PIR-based DFL methods still have a practical limitation that they usually require a high deployment density in multi-person scenarios. In this article, we propose a new PIR-based DFL method which requires a significantly lower deployment density in the multi-person scenario. The proposed method is based on deep learning. Especially, instead of directly utilizing the commonly used neural network architectures, we proposed a task-specific network architecture that integrates the domain knowledge of PIR sensors to improve its localization performance. Specifically, following the idea of modular learning, we design the proposed network architecture to contain two modules for counting the number of existing persons and another is for predicting their locations, respectively. Moreover, we also design the modules for person counting and localization to contain two secondary modules for different sub-tasks. Besides, to further improve the localization performance, we also propose two data augmentation strategies that aim at enhancing the training data diversity from the aspects of the moving speed of existing persons and the influence of surrounding noise. Through the proposed method, we succeed to remarkably reduce the deployment density of the traditional PIR-based methods by about 76%, while maintaining the localization accuracy.