Driving maneuver detection using knowledge distillation networks
Abstract
In this thesis, we examine the current state of Advanced Driving Assistance Systems (ADAS) and their relation to maneuver prediction in the literature. We then attempt to solve the problem of variable inter-driver behavior by applying a novel distillation learning system using RoadLab data on tracked driver cephalo-ocular gaze behavior in tandem with high-resolution CANbus data. Current training-based methods in maneuver prediction are potentially subject to underfitting as drivers may exhibit different behavior when preparing to maneuver, but it has been shown that drivers can be grouped into at least two distinct behavior models. We use this information to personalize a deep neural network ensemble by distilling knowledge from a larger "teacher" network to a smaller "student" network. We change the networks' input data to a subset of that data during training. Various groupings of driving sequence data are tested for prediction accuracy within this system, particularly against a validation driving sequence belonging to a specific driver group.