Indoor localization+
There is no doubt that Indoor Localization is indispensable to many service applications ranging from hospitals, malls, and even parking lots to name a few [8, 21, 43-45]. For example, indoor navigation is helpful for the customer to find the path to their desired shops in a large shopping mall. In...
Saved in:
Main Author: | |
---|---|
Other Authors: | |
Format: | Final Year Project |
Language: | English |
Published: |
Nanyang Technological University
2021
|
Subjects: | |
Online Access: | https://hdl.handle.net/10356/153130 |
Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Institution: | Nanyang Technological University |
Language: | English |
Summary: | There is no doubt that Indoor Localization is indispensable to many service applications ranging from hospitals, malls, and even parking lots to name a few [8, 21, 43-45]. For example, indoor navigation is helpful for the customer to find the path to their desired shops in a large shopping mall. In hospital buildings, patients with dementia can easily be located and taken care of through such systems [46].
In underground parking, these systems can help users locate their vehicles.
Indoor Localization has attracted many researchers’ attention over the past few years and continues to remain an active area of research. There has been a noticeable shift in the field of indoor localization,
from using extra hardware to more reliable robust techniques using existing wireless technologies coupled with machine learning. As a result, fingerprinting based on WiFi RSS signals has been largely adopted due to its effectiveness, simplicity, and minimal additional hardware requirements. However, its positioning accuracy is drastically reduced as a result of insufficient training data, as well as arbitrary fluctuations in RSS signals caused due to the multi-path phenomena as well as fading. This
report presents using an effective data augmentation scheme involving the use of Generative Adversarial Networks (GAN’s) to improve the performance of the model by synthetically producing additional training samples without the need for any extra human effort. Our proposed technique GAN+, has significantly improved the accuracy and efficiency of the entire positioning system. GAN+ is a workflow that uses Dirichlet distribution coupled with GAN’s to generate augmented data. The proposed technique was tested using a popular multi-building multi-floor dataset: UJIIndoorLoc. The
technique was further tested using another multi-floor dataset – N4. The N4 dataset was generated at the School of Computer Science and Engineering building located at Nanyang Technological University, Singapore.
GAN+ further makes use of Convolutional-Neural-Networks (CNN’s) time-series based neural networks in order to determine the user's location. CNN's have been used to leverage temporal dependency between time series RSS readings producing superior results as compared to a simple Deep Neural Network (DNN). Using multiple consecutive RSS readings reduces both the noise and randomness present in each separate RSS reading thereby enhancing the localization performance.
Experimental results show that the dataset generated by GAN+ can achieve an average location error that is more than ten times lower than that of the original dataset. Therefore, the proposed data augmentation scheme validates the feasibility of using GAN’s in the domain of indoor localization.
GAN+ trains a separate pair of generator and discriminator for each unique location to produce new synthetic samples for that particular location. While the augmented dataset generated does increase performance drastically, it is a computationally expensive and time-intensive task. This results in a bottleneck for the development of practical implementation of GANs. The use of transfer learning is thus proposed to initialize both the generator and discriminator neural networks for each unique
location. The transfer learning technique proposed significantly reduces training time. Experiments with both fine tuning and feature extraction are performed in order to observe the impact on model performance and training time. The results show that there is less than 0.1% drop in accuracy when using Transfer Learning. |
---|