Infrastructure-less based positioning:Localization in GNSS-denied environments
Islam, Saiful (2019)
Islam, Saiful
2019
Tietotekniikan DI-ohjelma - Degree Programme in Information Technology
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2019-08-30
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-201908263020
https://urn.fi/URN:NBN:fi:tuni-201908263020
Tiivistelmä
Research on the uses of navigation and positioning services is considered to be stable and such uses have an exponential growth in outdoor environments. The satellite-based positioning systems provide us an affordable outdoor positioning service and it becomes part of our daily lives. However, all the benefits from the satellite systems may be lost when we enter the indoor spaces or such places where satellite signals are not received due to high attenuations of walls and floor materials. Many users such as rescuers, mine workers, firefighters maneuver have their activities in satellite-denied area and need to work without past information about their surroundings. That is why we need some other positioning services, which can aid us in satellite denied areas. Different technologies can be used for indoor positioning. However, the use of indoor positioning systems may involve extra infrastructure and setup, making the indoor positioning system more complicated and costlier than the satellite-based outdoor technology. An enhanced positioning technology with ubiquitous coverage can address these issues that reduce infrastructure dependence. In this thesis, an infrastructure-less based positioning algorithm is studied. This algorithm relies on the aroma fingerprints of any closed areas. Ion mobility-based electronic noses (eNoses) were used here to obtain aroma fingerprints from different locations. The performance of eNoses in the case of location estimation in an indoor environment has been shown in the research. The data used in this research was collected from seven different locations at Tampere University (TAU) campus. Data in empty and crowded spaces were collected for each location for a total of about 600 s. A supervised machine-learning algorithm was used to process and estimate the probabilistic location. The non-parametric fingerprinting techniques were applied to determine the location from the measurements. The non-parametric system trained with a dataset containing location information of known places called the offline phase. In the online phase, real-time data from the unknown places were recorded and matched with the existing dataset to estimate user location. Five different classifiers were studied in the thesis to predict the location of a user. Using the Scikit-Learn library of Python, a software-based model was developed to evaluate different parameters and outcomes of different classifiers. The popular classifier is k -nearest neighbor (kNN), which correctly predicted about 38 percent of the locations. The impact of different distance metrics and the number of closest neighbors in the localization accuracy were also explored in the thesis. In addition, dimensionality reduction techniques were also applied to reduce correlations between the different electrodes as well as to reduce computational time and complexity. From the results, it is observed that when Principal Component Analysis (PCA) was implemented, the support vector machine anticipated an even more satisfying outcome from the kNN. On the other side, PCA selects eigenvectors that have more variances and removes those with fewer variances that sometimes do not fit with other classifiers. In terms of accuracy, the potential result was achieved from an unusual classifier called Stochastic Gradient Descent (SGD). SGD classifier estimated an object's location correctly up to 53 percent times under certain conditions. Also, some additional experiments were conducted by training the model with different environments. Classifiers are noted to be able to obtain improved accuracy when the model has sufficient environmental data. For example, when the model is well trained, the Random Forest Classifier (RFC) delivers more accurate results than the lightly trained model. On the other hand, RFC delivers poorly accurate results when the model is lightly trained. The impact of the data size on the accuracy was also studied in the thesis. The final experiment demonstrates that most classifiers perform well when the size of the training data is large compared to the size of the test data. However, even with small training data, the kNN classifier with Euclidean distance performed better under all circumstances.
It is hard to state which classifier is the winner from the experiments, but the highest predictive accuracy was achieved by the SGD classifier. SGD output, however, is not stable over time, and with the increase in test size, accuracy decreases. Instead, the stable performance of the kNN classifier with Euclidean distance makes it more reliable to use in any conditions. Also, there were some problems mentioned in the thesis before one could recommend using aroma fingerprints as a trustworthy method of localization.
It is hard to state which classifier is the winner from the experiments, but the highest predictive accuracy was achieved by the SGD classifier. SGD output, however, is not stable over time, and with the increase in test size, accuracy decreases. Instead, the stable performance of the kNN classifier with Euclidean distance makes it more reliable to use in any conditions. Also, there were some problems mentioned in the thesis before one could recommend using aroma fingerprints as a trustworthy method of localization.