Aprendizagem de máquina para sistemas de localização em ambientes internos RFID

Abstract

Object location in indoor environments is a challenging task due to the absence of physical contact or field of vision, reflective materials, excessive obstacles, and a high density of items. Several surveys propose using Radio Frequency Identification (RFID) technology and Machine Learning (ML) methods to construct Indoor Positioning System (IPS). However, using an object as a target class makes learning and prediction slow in large-scale environments. Another difficulty frequently encountered in the use of machine learning methods is the need to train all the positions of the tags mapped in the environment. This training increases the complexity of the system deployment process. To circumvent such problems, we present in this thesis four new approaches to location systems for indoor environments using RFID technology and machine learning techniques. The first approach uses hierarchical classification by dividing the environment into regions to reduce the number of classes in the training and prediction of the classifiers. The second approach applies the sliding window concept to statically extract new features from the Received Signal Strength Indicator (RSSI) attribute and then trains the Random Forest regressor algorithm using reference tags to predict the coordinates of the target tags. In addition to using the concepts of sliding windows and reference labels, the third approach also applies hierarchical classification. The fourth approach applies a genetic algorithm to identify which reference tags contribute most to the system’s accuracy. One of the main results got was a precision of 4.82 cm in an environment with 7,000 target tags, reducing by up to 92.91% the average error w.r.t. other IPS approaches presented in the literature, including more complex indoor scenarios. Another main contribution of the present thesis is the public availability of an RFID dataset for future research in the area of indoor location.