Degree
Doctor of Engineering (DEng)
Department
Electrical Engineering
Document Type
Dissertation
Abstract
In this dissertation, we propose a novel simulation-based device-free indoor localization and tracking system using the received signal strength indicators (RSSIs) of WiFi signals as the input features. The Feko channel-propagation simulation software is used to process the RSSI maps of the given arbitrary indoor geometry. In order to learn the dynamic information of high-dimensional RSSI time-series, we propose three procedures for the localization and dynamic tracking system.
First, The indoor geometry is partitioned into several equi-size zones and the localization problem is treated as the typical \multi-classification" problem. The advanced machine-learning techniques such as decision tree (DT) classifier, random forest (RF) classifier, and gradient boosting decision tree (GBDT) classifier are employed to predict the zone-index sequence based on the RSSI features for instantaneous localization. However, the advanced machine-learning techniques fail to infer the dynamic information inherent in the trajectory of the moving object. Thus, we then propose the previously mentioned advanced machine-learning techniques in conjunction with the Viterbi algorithm over hidden Markov models (HMMs) for dynamic tracking system. Since we have partitioned the indoor geometry into several equal-size zones, the proposed dynamic tracking approach suffers from the widely known rasterization effect due to the restricted zoning resolution. To mitigate the rasterization effect, we finally propose both Kalman filter and Kalman smoother in conjunction with the GBDT classifier for trajectory prediction and regularization, respectively.
The simulation results demonstrate that the GBDT classifier reach the average Euclidean-distance errors of 3.392m and 3.562m over eight zones and sixteen zones respectively, whereas the GBDT classifier in conjunction with the hidden Markov models reach the average Euclidean-distance error of 2.008m and 1.639m over eight zones and sixteen zones, respectively. Furthermore, the Kalman filter and the Kalman smoother in conjunction with the GBDT classifier reach the average Euclidean-distance error of 1.58m and 1.33m, respectively. The computational complexity of the proposed indoor localization and dynamic tracking system is also proposed in this work. As a result, the proposed novel indoor localization and tracking scheme would be very promising and convenient for many devices-free indoor-localization applications such as smart health-care, security surveillance, and first-responder tracking in the future.
Date
7-8-2023
Recommended Citation
Liu, Guannan, "Learning Dynamic Information of High-dimensional Signal Time-series using Advanced Machine Learning/Artificial Intelligence" (2023). LSU Doctoral Dissertations. 6228.
https://repository.lsu.edu/gradschool_dissertations/6228
Committee Chair
Wu, Hsiao-Chun