Breaking the Trade-Off: High-Fidelity Indoor Localization Using Intra-Model Distilled 3D CNNs

Reliable indoor localization is essential for smart environments but is undermined by fluctuations in WIFI received signal strength due to multipath propagation, shadowing, and fading. Conventional 2D convolutional networks struggle to effectively capture these temporal variations, while 3D CNNs, though capable of modeling such dynamics, introduce prohibitive computational overhead, including increased memory footprint, inference latency, CPU/GPU cycles, and energy consumption, which hinders their practicality on edge and mobile platforms. To address this gap, a lightweight 3D CNN framework integrated with intra-model knowledge distillation (IMKD) is proposed. Unlike traditional knowledge distillation methods that rely on large teacher–student paradigms, IMKD employs a self-contained strategy in which the deepest classifier of the network serves as an internal teacher, progressively guiding shallower classifiers. This mechanism strengthens temporal feature extraction, improves optimization stability, and reduces computational cost without compromising accuracy. Numerical evaluations on the UJILIB dataset reveal that the proposed framework achieves comparable accuracy and average positioning error to the vanilla CNN and superior performance to the lightweight CNN, with an 84% reduction in computational cost in terms of floating-point operations compared to the vanilla model, underscoring its strong potential for real-time edge deployment. Ablation studies further reveal how the synergistic effect of lightweight 3D modeling and intra-model guidance offers valuable design insight for future deployment in resource-constrained environments. By combining robust temporal modeling with an efficient self-distillation strategy, the proposed study advances indoor localization beyond current paradigms and provides a practical, deployable framework for fast and accurate localization in next-generation smart environments.