Deep learning-driven indoor positioning using WiFi fingerprints

Indoor Positioning Systems (IPS) extend the functionality of the Global Positioning System (GPS) to indoor environments where GPS signals are usually unavailable, significantly enhancing the utility, safety, and efficiency of indoor spaces. These systems are crucial for navigation in complex environments such as airports and museums, for indoor evacuation during emergencies, and for improving rescue response times. With the widespread deployment of WiFi and the proliferation of smart devices, WiFi fingerprint-based indoor positioning has become one of the most practical methods for positioning indoor mobile users. However, the Received Signal Strength Indicator (RSSI) of WiFi is highly sensitive to indoor environments; movements of people, multipath effects, and changes in temperature and humidity can affect RSSI values, thereby degrading positioning performance. This thesis introduces DNpos, an advanced deep learning model for noise-resistant multi-building and multi-floor indoor positioning. DNpos integrates a Denoising Autoencoder (DAE) based on a Deep Neural Network (DNN) to counteract environmental noise and employs an attention mechanism to focus on the most informative features, optimizing the use of computational resources. It utilizes a Convolutional Neural Network (CNN) for local perception of RSSI data features and a DNN to output the position. Additionally, DNpos introduces a Gaussian Mixture Model (GMM)-based algorithm for extracting representative training and validation sets, addressing positioning fluctuations in small datasets and overcoming the limitations of random selection methods. The performance of DNpos was validated using standard datasets, demonstrating outstanding noise resistance and generalization performance. The system achieved high success rates for building-level and floor-level positioning and also showed excellent performance in coordinate positioning, significantly outperforming current leading technologies.