Deep Learning Approach to Identify Locations from Speakers Emitting the Same Ultrasound Signal in an Indoor Positioning System
Abstract: An Indoor Positioning System (IPS) utilising ultrasound technology over a single chan- nel based Fingerprinting method poses significant challenges to achieve high level of accuracy. In recent years, the deep learning techniques for IPS has achieved consider- able performance over the traditional machine learning techniques as per the research community. The thesis explores the accuracy of deep learning methods for an ultrasound based Indoor Positions System developed by the DEJ Technology GmbH which is also known as Koopango IPS. The research focuses on the answers to the question: How deep learning techniques can improve accuracy in predicting location coordinates in an indoor environment using ultrasound based Fingerprinting Indoor Positioning System? This paper provides an in-depth functionality of the Koopango IPS, an overview of how data is collected and processed exploiting the concepts of audio processing, and examines deep learning methods to improve accuracy of IPS for localization of (x,y) coordinates in an indoor space using a Design Science research methodology. The comprehensive data-sets consists of audio signal recordings with corresponding location coordinates in two distinct indoor environments are used. The first data-set is used for training and evaluation of the model, whereas the second data-set evaluates the behaviour of the same model in a different environment. The solution obtained provides a significant performance of the deep learning model with a mean squared error of 2.67 meter square and 0.55 meter square in the predicting the location (x,y) coordinates. The result represents a significant improvement of accuracy over the existing Koopango IPS using single channel playing same ultrasound signal. The thesis is concluded with the observations and the possibility of incorporating factors like reflection and reverberation in the future research.
The repository showcase the python notebook with data analysis and Deep Learning model designed to improve accuracy of the Koopango IPS