Master of Science
Bauer, Michael A.
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.
Summary for Lay Audience
Advanced Driving Assistance Systems (ADAS) are systems implemented in vehicles composed of computer and sensory equipment that augment the driver's natural abilities. These systems may provide indicators such as signals or additional data feeds. Neural networks are a collection of mathematical operations that are applied in sequence to an input. They are distinct from simple equations in that they can modify their own equation coefficients to try to mimic a desired output. Neural networks modify their own equations using many examples. Once the equation predictions are close enough to a desired output, they can be used with real-world inputs in place of examples to make accurate predictions. In this thesis, we evaluate the use of distillation neural networks as a tool in ADAS. Particularly, we are using distillation networks to predict driver maneuvers. A driver maneuver may be a left turn, a straight driving sequence, or a right turn. If we represent these maneuvers as numbers, they can be used as example desired outputs for a neural network. As example inputs, we use the driver's eye movement and some sensors (i.e. CANbus sensors, a standard sensor protocol) augmenting the vehicle. A distillation network is a combination of multiple neural networks, where one network acts as a teacher, and the other acts as a student. When making predictions, the student factors the teacher's predictions into its decisions. We show that a well-trained student network works better for maneuver prediction than just a teacher network alone.
Windsor, Kyle, "Driving maneuver detection using knowledge distillation networks" (2020). Electronic Thesis and Dissertation Repository. 6844.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.