Position Quantization Approach with Multi-class Classification for Wi-Fi Indoor Positioning System

Abstract

Indoor positioning system is a challenging problem due to the variety of environment and unreliable of data that are used for a prediction of the position. For Wi-Fi based indoor positioning system, signal intensity used to predict the coordinate of the device are known to fluctuate greatly despite being measured at the same position. Therefore, significant errors are often found when solving this problem with regression algorithms. A quantization of co-ordinate data into position IDs can mitigate the fluctuated noises in the data and is able to reformulate the problem into a multi-class classification problem. The error in positioning can then be computed from the distance between the true co-ordinate and the predicted co-ordinate. The experiment shows that Random forest classification can predict the position with the error in positing at 5.65 meters on average when the quantization is applied with threshold setting to 1 meter.