Neural network optimization for indoor person localisation using capacitive sensors

Nowadays one of the main aspect that can be analyzed in smart home society is human localisation and movement estimation. It can lead to a better quality of life with reduced costs for example by handling other existent smart technologies as automatic heating and lightning. Moreover, indoor movement monitoring can be really useful in assisted-living for elderly people. This kind of system must be privacy aware and easy to use and a tagless passive positioning system can be a good choice. Among the different type of passive sensing one of the most interesting is the capacitive sensing. The need of cheap hardware and the possibility to analyse the acquired data by a microcontroller make capacitive sensors easy to produce and to prototype. This sensors are the main part of a project in progress "Capacitive sensing for indoor human localisation" that is related to the thesis work. Among the different modes of working of capacitive sensing, the loading mode has been chosen since a single electrode is used. The principle of working is simple: as the person gets closer to the sensor, the capacitive coupling between body and sensor increases, by measuring this capacitance, the distance from the sensor can then be inferred. Hybrid technology has been explored: the position of the person within the room has been monitored using two systems. The first uses four capacitive sensors , while the second, used as reference because of its accuracy, is based on four ultrasound anchors that can localize a mobile tag. The reference position and the capacitive sensors readings have been collected concurrently: the former data has been used for training data labelling and inference testing, the latter data has been pre-processed with digital filtering, then neural networks has been used to infer the correct position. The obtained data set has been used in the thesis work that will focus on different Machine Learning approaches to localise and evaluate movement of a person in a room. Different Neural Networks have been trained and tested to report location estimation as a pair of X/Y coordinates and performance are evaluated in terms of mean squared error (MSE) and euclidean distance error (EDE). However, also training and validation error curves obtained during the training, validation and test phases have been monitored, in order to prevent overfitting or underfitting. Three main types of Neural Networks have been analysed. Fully connected Neural Networks: made by a series of fully connected layers, also known as Multilayer perceptrons, used as staring point to better understand the general behaviour of a neural network 1D-Convolutional Neural Networks: used to derive features from short segments of the overall data set. This can be useful to analyse time sequences of sensor data and infer the movement from collected data. Long Short Time Memory (LSTM): artificial recurrent neural network (RNN) architecture able to process single data points, but also sequences of data, this applies well to the movement estimation of a person in a room