MCLS-AI: Bridging Activity Identification and Multipath RFID Signals via Contrastive Learning

Radio Frequency Identification (RFID) technology is increasingly utilized in Internet-of-Things (IoT) applications for object localization and activity identification, offering advantages such as low cost, compactness, and no battery dependency. Despite its potential, RFID's effectiveness in real-world scenarios is hindered by multipath interference from various indoor elements, which complicates signal interpretation and accuracy. Traditional deep learning models, such as CNNs, struggle with these complexities, especially in environments lacking extensive, labeled datasets. We propose MCLS-AI (Multipath-Aware Contrastive Language-Signal Object Activity Identification), a novel framework leveraging contrastive learning to address these challenges. By employing a decoupling mechanism that analyzes the pseduospectrum within raw signal mixes, MCLS-AI effectively isolates and processes the distinct paths of RFID signals. Combined with textual activity labels, this approach enhances accuracy in identifying object activities. Our results demonstrate that MCLS-AI outperforms both deep learning models like ResNet-18 and YOLOv8, as well as conventional methods like Neural Net, achieving the highest accuracy and F1-Score, showcasing its superior effectiveness.