Date of Award
5-14-2024
Degree Name
Doctor of Philosophy (PhD)
Degree Type
Dissertation
Abstract
In domains like automation, particularly in advanced driver-assistance and collision avoidance systems for vehicles, the need for reliable sensing and predictive capabilities is paramount. While current sensor technologies allow for quick responses to safety-critical events, there is a growing emphasis on predictive capabilities to anticipate and prevent such incidents. This opens new ideas for the development of sophisticated RF-sensors and algorithms capable of accurately predicting potential harms before they occur. Moreover, along with the benefits of enhanced sensing capabilities comes the imperative to safeguard user-privacy. Various techniques, including encryption and differential privacy, are employed to ensure that RF-based services do not compromise user data, while innovative approaches such as discarding irrelevant data and utilizing privacy-preserving measurements like received signal strength and Doppler shift offer promising avenues for balancing functionality with privacy concerns. In this thesis, we integrate the power of radio frequency sensing and rapid progress of the data-driven learning to bring forth ideas that can be applied to safety-critical applications for enhancing efficiency and user experience in an increasingly interconnected world. We present three products that are built using three types of radio devices with differing in their operating principles and thus, their capabilities to be utilized in three unique scenarios; network-based cooperative sensing, standalone sensing, and sensing as a surveyor/monitor. Our first novel contribution is an infrastructure-free approach to collision prediction using ultra-wideband (UWB) signals and inertial sensing. They employ a cooperative strategy based on pairwise ranges and velocities to predict future collisions, utilizing an improved algorithm to estimate relative kinematics despite noisy measurements. This method is complementary to existing technologies dependent on object properties, with UWB chosen for its precise measurements and independence from indoor object properties. We continue our endeavor by introducing a systematic shift in the type of sensor used, by instead using a standalone sensor, such as radar, for collision prediction in noisy, cluttered environments with dynamic motion. We utilize the radar-Doppler data and a convolutional neural network (CNN) to predict collisions, adapting features from the environment to handle inaccuracies. Online learning and automated labeling techniques are employed to make the CNN adaptable, with experiments resulting in a labeled dataset for validation against other methods. Finally, in order to provide a method to implement these safety-critical applications, we investigate how the sensors can `police' or survey an area and build applications from extracting only the relevant data from the RF-signal measurements through a new measurement called Doppler spread. Outdoor experiments in a densely populated area are conducted, generating a labeled database and examining Doppler spread's effectiveness for a privacy-preserving localization system based on fingerprinting.
Language
English (en)
Chair
Neal Patwari