Balancing of associated attributes for a viable indoor real-time location system with obstruction detection

Abstract

Outdoor location determination is often achieved by using GPS, but indoor location determination is not possible with GPS due to the limited link to satellites from indoor environments. Research has indicated that indoor location determination is applicable in a variety of domains, including asset location, location of people, emergency evacuation, participant attendance in a venue, tracking of people’s mobility, and obstruction location. Cost, energy consumption, interference, coverage, detection range and form factor are some of the constraints reported in the literature. The attributes for further research were derived from these constraints. Bluetooth Low Energy was identified as the most suitable technology with which to design a model that would achieve an optimal balance between the identified attributes. The research used Unified Modelling Language to document the model and design science methodology to design, test and validate the model. The introduction of obstructions in the path of transmission often affects the received signals and hence affects the location determination. The connection quality indicator was used in this model to determine location instead of the widely used fingerprint method, whose data becomes unusable over time as it becomes stale and inaccurate. The design was tested with a variety of obstructions, including drywall partitions, glass, solid brick walls, metal sheets and Perspex, all of which were utilised to resemble a typical office environment. The received signal strength indicator measurements from low power nodes were filtered and smoothed using mean, median and mode statistics. This received signal strength indicator data was then used by support vector machine, k-nearest neighbour and artificial neural network machine learning models to determine the location and impact of the obstructions in the path of Bluetooth Low Energy transmission. The results obtained from machine learning and prediction revealed that the location of obstructions was determined to be within an acceptable level of accuracy. In particular, knearest neighbour performed the best compared to support vector machine and artificial neural network using the mean squared error, mean absolute error, root mean square error and Rsquared score metrics. In particular, mean squared error and mean absolute error metrics revealed the best results. The study indicates that machine learning can therefore be used to determine positions of semi-fixed obstructions within a select indoor environment.