Indoor Localization and Sensing Techniques for Millimeter-wave Wireless Systems

Millimeter-wave (mmWave) frequencies, characterized by short wavelengths, large available bandwidth, and highly directional communication, hold immense promise for not just delivering multi-Gbit/s data rates, but also for enhancing the accuracy and efficiency of device-based localization and device-free sensing capabilities in indoor environments. The quasi-optical propagation of mmWave signals enables high-accuracy localization algorithms that employ geometric approaches or machine learning models. However, these algorithms require the knowledge of the environment, may entail a burdensome training data collection process, or are computationally infeasible for implementation on resource-constrained commercial off-the-shelf (COTS) devices. mmWave radars enable device-free sensing to accurately sense subjects of interest indoors. However, commercial radars have very short range and are susceptible to occlusion events. Hence, dense deployments in large indoor spaces are inevitable. This raises an important challenge of calibrating the radar network, i.e., estimating the location and orientation of the radars with respect to a common reference radar, in order to collectively fuse information from them. This thesis substantially contributes towards designing and evaluating lightweight algorithms, in an attempt to solve the above-mentioned challenges along two research lines. In the first part of the thesis, we propose to localize a mmWave client device using tiny neural network (NN) models. These NNs have fewer parameters (e.g., the number of neurons and the number of hidden layers), and are computationally lightweight, making it easier to implement them on resource-constrained COTS devices. To relieve training data collection efforts, we resort to a self-supervised approach by bootstrapping the training of our NN through location estimates obtained from a state-of-the-art localization algorithm. Our proposed idea is extensively evaluated via simulations, outperforming the existing state-of-the-art geometry-based localization schemes. In order to experimentally validate our proposed tiny NN-based localization scheme, we exploit the channel measurements obtained from an indoor 60-GHz double-directional channel sounder. We first process the raw channel measurements to extract multipath components (MPCs) and their corresponding signal parameters, and propose an algorithm to recursively cluster the azimuth and elevation angle of arrivals (AoAs) measurements. This helps us extract the dominant MPCs and their corresponding AoAs, so as to be compatible with our NN model. Extensive evaluation shows that our lightweight models can accurately localize a device in the indoor space, thereby validating our simulation results. As the indoor environment scales up in size, the tiny NN will require a lot more training data and more parameters to accurately localize a client device. Thus, instead of training and deploying a single deep model, we train multiple tiny NNs, one for each smaller subsection of the indoor space. In order to decide the best NN autonomously, we propose two NN-switching schemes: one based on the innovation measured by a Kalman filter, and the other based on the statistical distribution of the training data. This allows us to localize a client device in a distributed manner. In the second part of the thesis, we focus on device-free sensing using mmWave radar networks. Existing state-of-the-art indoor sensing applications employ a single mmWave radar. However, as mmWaves (especially from 57 to 81 GHz) can be easily blocked by obstacles, including humans, the use of multiple radars, i.e. a network of radars, is inevitable. This poses the fundamental challenge of self-calibrating the radars to a common reference system, so as to fuse the information from them. Hence, we propose mmSCALE, a self-calibration system that automatically estimates the location and orientation of the radars with respect to a reference radar. Note that self calibration problem in radar networks is inherently a passive localization problem. We emphasize that mmSCALE requires no specific target trajectories or controlled conditions for calibration, autonomously assesses the calibration quality over time, and is robust to occlusion events (in presence of multiple subjects). In order to fuse data from multiple radars, we extend our system and propose ORACLE, an autonomous plug-and-play system, that automatically calibrates the radars in the network and fuses the tracking information of the people tracked by different radars at a fusion center, thus enhancing the resilience of the subject localization process in case of occlusions. While client device localization is an active localization problem (relying on the mmWave signal parameters obtained from the initial link establishment between the client and the access point (AP)), the calibration of a network of radars is a passive localization problem (relying purely on the signals reflected off the environment). Our proposed schemes are lightweight and substantiated by experimental evaluation, thus rendering them amenable for implementation and deployment on resource-constrained COTS devices.