Lately, models utilizing channel impulse response (CIR) data for training deep neural networks used for positioning in radio networks have shown promise, particularly in simulated indoor environments. Research has extended to real outdoor setups as well. In this study, deep neural networks originally designed for image classification were employed to estimate positions using both real and simulated outdoor CIR data. A ray tracing simulator was utilized to generate a simulated dataset which corresponded to a real-world dataset. Models were trained and tested on both datasets. To facilitate training on simulated data and testing on real data, a generative adversarial network (GAN) was employed. The thesis concludes that deep neural networks can effectively be used in real outdoor scenarios, but dense data sampling is likely necessary to achieve satisfactory performance across an area. Additionally, it was found that the simulated data used in this study differed significantly from reality, and the employed GAN could not effectively bridge this gap. Consequently, models trained on simulated data performed poorly when tested on real data. However, it was found that deep neural networks significantly outperformed the baseline K-nearest neighbor algorithm when trained and tested on simulated data. However, this was the only case where such a significant advantage for the deep models was observed.
| Identifer | oai:union.ndltd.org:UPSALLA1/oai:DiVA.org:liu-205237 |
| Date | January 2024 |
| Creators | Engström, Gunnar |
| Publisher | Linköpings universitet, Kommunikationssystem |
| Source Sets | DiVA Archive at Upsalla University |
| Language | English |
| Detected Language | English |
| Type | Student thesis, info:eu-repo/semantics/bachelorThesis, text |
| Format | application/pdf |
| Rights | info:eu-repo/semantics/openAccess |
Page generated in 0.0025 seconds