Nowadays, there is a growing demand of location based services for indoor environments. Unfortunately, the commonly used Global Positioning System (GPS) is inefficient in most indoor environments. Since, several indoor localization methods have been investigated to address this problem. Among these methods, methods based on the receiver signal strength have been widely investigated due to their simplicity and easy implementation. Specifically, the receiver signal strength indicator (RSSI) based method known as fingerprinting localization approach has attracted significant interest due to its performance at low cost. This method does not require a specific hardware and network infrastructure. However, due to the RSSI fluctuation, most fingerprinting localization approaches do not achieve good performance as expected.
This thesis first examines the performance of some artificial neural network based fingerprinting localization methods. In this part, we proposed a multiple neural network based fingerprinting localization model whose performance has been compared to those of a K-nearer neighbor (KNN) and a deep neural network based fingerprinting localization approach that exploits deep auto-encoder method to extract feature data. In the second part, we proposed a wavelet scattering transform based feature extraction method to improve the performance of deep neural network based fingerprinting method. The merit of the wavelet scattering transform based feature extraction is that the wavelet scattering transform provides multi-scale information of the RSSI data.
Through different experiments, the proposed methods have achieved good localization performance compared to others state-of-the art existing fingerprinting approaches used in this study.