High-Resolution Two-Target Indoor Tracking Using Conv-TasNet on Ceiling-Mounted PIR Analog Signals

Abstract

Passive infrared (PIR) sensors are ideal for indoor positioning due to their low cost, low power consumption, and privacy preservation. However, existing PIR systems primarily target single-occupant scenarios; separating mixed analog signals from co-occurring individuals remains a critical challenge. This study introduces a novel framework utilizing Conv-TasNet—originally designed for audio source separation—to isolate overlapping PIR signals from two targets. To resolve postseparation permutation ambiguity, we implement a joint Mel-frequency cepstral coefficients (MFCC) and XGBoost classification model. Experimental evaluation using a ceiling-mounted PIR system (3.3 × 3.3 $\text{m}^{\mathbf{2}}$) demonstrates that the MFCC-XGBoost model achieves 73% tracking accuracy and a 0.39 m path error. This reduces localization error by 42% compared to state-of-the-art multitarget systems, while maintaining an inference latency of just 0.07 s. These results prove that advanced algorithms can achieve precise, real-time, two-target tracking using cost-effective hardware, offering a practical, and privacy-preserving localization solution.