A Hybrid CNN-Transformer Deep Learning Framework for Accurate Wi-Fi Indoor Positioning

Abstract

Accurate indoor positioning remains a significant challenge due to the unpredictable nature of indoor radio signal propagation. This study presents a novel Wi-Fi fingerprinting-based positioning system using a hybrid deep learning architecture that combines Convolutional Neural Networks (CNN) with Transformer encoders. Unlike traditional algorithms such as KNN, WKNN, SVR, and DeepFi, the proposed CNN-Transformer model leverages the spatial feature extraction capabilities of CNN and the global sequence learning strength of Transformers to enhance indoor positioning accuracy. A unique regression head is integrated to predict precise coordinates directly from raw RSSI input vectors. The proposed CNN-Transformer model outperformed all other algorithms with a Mean position error (MPE) of 1.76 m and a 95th percentile error of 3.2 m. Furthermore, the Cumulative Distribution Function (CDF) analysis revealed that 90% of predictions were within 2.8 m, demonstrating high accuracy and consistency. Although the model incurs higher inference and training times, the significant improvement in accuracy makes it suitable for real-time applications in complex indoor environments. These results underscore the effectiveness of combining CNN and Transformer architectures for robust and scalable indoor localization systems.