Abstract
Fingerprint-based indoor localization leveraging Wi-Fi and Bluetooth received signal strength stands as a prominent infrastructure-less positioning technique. While positioning accuracy is essential, achieving it under practical deployment constraints remains a key challenge. Convolutional Neural Networks (CNNs), recognized for their robust pattern recognition capabilities, hold the potential to significantly enhance indoor positioning accuracy. However, the inherent architectural complexity of deep CNNs restricts their deployment on resource-constrained devices. Knowledge distillation (KD) offers a viable strategy by transferring of knowledge from a complex model to a simpler, more efficient one. This study proposes a novel lightweight 2D CNN architecture integrating squeeze-and-excitation (SE) modules with an adaptive temperature guided iterative self-knowledge distillation (SKD). The SE mechanism dynamically recalibrates CNN filter responses, prioritizing those capturing salient features, while the iterative SKD progressively refines the model's performance during the training process. Unlike conventional KD approaches necessitating distinct teacher and student models, our proposed technique employs a single lightweight model, significantly reduces computational overhead. Empirical evaluation on the HDLC public datasets demonstrates that our architecture, without the incorporation of SKD, yields a notable 8.32% improvement in positioning accuracy over conventional CNNs, achieving a 3D average positioning error (APE) of 2.60 m. Furthermore, the integration of the iterative SKD strategy further enhances the positioning accuracy to 1.66%, resulting in a 3D APE of 2.24 m. These findings underscore the efficacy of the proposed framework as a resource-efficient and practical solution for accurate indoor localization, facilitating its real-time implementation on resource-constrained platforms.