Author ORCID Identifier:

https://orcid.org/0009-0008-9870-022X

Date of Graduation

5-2026

Document Type

Dissertation

Degree Name

Doctor of Philosophy in Computer Science (PhD)

Degree Level

Graduate

Department

Electrical Engineering and Computer Science

Advisor/Mentor

Gauch, John

Committee Member

Fredrick, David

Second Committee Member

Zhang, Lu

Third Committee Member

Li, Qinghua

Keywords

Neural Networks; Autonomous Perception, TinyBEV

Abstract

Autonomous perception systems must operate reliably under uncertainty arising from noisy observations, incomplete supervision, and hardware constraints. This dissertation investigates the design of efficient deep neural networks for autonomous perception through a unified perspective that treats uncertainty, efficiency, and sensing as interconnected challenges. The first contribution develops adaptive extensions of unbiased risk estimators, including eSURE and ePURE, enabling unsupervised training of deep neural networks for magnetic resonance image denoising under Gaussian and Poisson noise. However, these methods rely on known noise assumptions, which motivates the second contribution: a unified diffusion and Bayesian risk framework that estimates and adapts to unknown and mixed noise distributions, achieving robust denoising without prior knowledge of noise characteristics. Building on this foundation, the third contribution introduces TinyBEV, a compact camera-only Bird's-Eye-View perception framework that distills knowledge from a large multi-modal teacher model. TinyBEV achieves approximately five times faster inference and reduces parameters by over 70% while maintaining strong performance across detection, mapping, forecasting, and planning tasks. Finally, this dissertation presents a task-driven optics-sensor-model co-design framework that jointly optimizes image formation and semantic segmentation. Experimental results demonstrate improved robustness under noise, blur, and quantization, achieving higher mIoU and better performance on challenging classes. Collectively, these contributions establish a unified framework for learning under uncertainty and constraints, enabling efficient, robust, and deployable perception systems for real-world autonomous applications.

Download

Share

COinS