ADCLoc: A robust and adaptive CSI-based device-free passive indoor localization approach for dynamic environments

The proliferation of emerging sensor technologies and the expansion of the Internet of Things (IoT) have sparked significant interest in advanced indoor localization methodologies. Device-free passive localization through Channel State Information (CSI) fingerprinting has attracted considerable research attention, as it eliminates the need for target participation, rendering it particularly valuable for location-based IoT applications. However, ensuring accuracy in dynamic real-world environments remains challenging due to the inherent trade-off between deployment costs and localization precision. Variations in CSI caused by environmental fluctuations can degrade system performance, underscoring the necessity for robust and adaptable solutions. In this study, we propose ADCLoc, a novel CSI-based device-free passive localization framework specifically designed for dynamic environments. ADCLoc employs a fusion model that integrates spatio-temporal redundancies in CSI amplitude and phase data, thereby enhancing the discriminative capacity of localization fingerprints. Central to ADCLoc is an adaptive convolutional neural network (AdaptCNN) incorporating a meta-learning dual-stream architecture for unsupervised domain adaptation. This design enables continuous adaptation to fluctuating CSI conditions inherent in dynamic environments while maintaining high performance without requiring extensive retraining. A six-day evaluation under controlled environmental modifications—including furniture rearrangement and door/obstacle configuration changes—demonstrates that ADCLoc surpasses existing methods in both localization accuracy and robustness.