Doctoral Dissertation Oral Proposal, Mahan Tabatabaie

Title: Ubiquitous Driver Behavioral Sensing and Learning
Ph.D. Candidate: Mahan Tabatabaie
Major Advisor: Dr. Suining He
Associate Advisors: Dr. Bing Wang and Dr. Song Han
Date/Time: Thursday, May 2nd, 2024, 2:00 pm
Location: HBL1947

Meeting link:
https://uconn-cmr.webex.com/uconn-cmr/j.php?MTID=m92c20047b9dcb3dcbf967aafabbba8c6
Meeting Access Code: 2634 470 4049
Password: 1234

Abstract:

In this thesis proposal, we aim at investigating ubiquitous driver behavioral sensing and learning techniques in terms of (1) coarse-grained driving behavioral patterns, such as driving routines at different times of the day or regions within the city, and (2) fine-grained driving actions, such as the left and right turn maneuvers. Such driving behavioral sensing and learning approaches can support various practical applications, including prevention of unauthorized driving, enhancement of driving safety, and improved connected autonomous vehicles through timely and potentially useful feedback.

This thesis proposal will focus on approaching the following two challenges of developing driver behavioral sensing and learning systems. First, due to complex spatio-temporal settings and drivers’ decision-making behavior patterns, different drivers may potentially have very similar active regions making differentiating them based on simple features extracted from their historical GPS trajectories challenging. Second, due to the varied or altered deployment settings and environments, such as different drivers, sensors, or road environments, in fine-grained behavioral analytics based on inertial measurement units, the existing deep learning-based models may not adapt well in practice. For instance, urban or rural road environments and the subsequent heterogeneous sensor measurements may affect the representations within the driver maneuver data.

To overcome above technical challenges and develop ubiquitous driving behavioral sensing and learning systems, we will first investigate extracting spatio-temporal driver behavioral patterns through fusing features with the information from various contextual factors and points-of-interest to create distinct mobility fingerprints for different drivers. Second, to learn the dynamic and heterogeneous driver maneuvers for fine-grained behavior identification, we will design a novel and adaptive driver maneuver identification approach based on multi-representation learning and meta model update for extracting feature relations from the drivers’ maneuvers.