Enhancing Autonomous Vehicle Resilience Through Engineering Requirements and Snow-Adaptive Lane Detection
Abstract
This thesis investigates two distinct but interrelated challenges in the development of resilient autonomous vehicle (AV) systems: the formalization of engineering requirements for AV perception subsystems and the enhancement of visual lane detection under snow-covered road conditions. In the first study, field experiments were conducted using a campus-deployed autonomous research vehicle to evaluate the impacts of perception related failures including GPS outages, HD map inconsistencies, and weather interference—on vehicle operation. These findings were used to develop a set of qualitative engineering requirements that promote AV resilience through proactive design. In the second study, a custom snow-focused lane detection dataset was developed and used to evaluate the Cross-Layer Refinement Network (CLRNet) across three backbones: ResNet-18, ResNet-34, and ResNet-101. Results showed that all models achieved strong performance (F1 = 0.90), with detection accuracy declining as snow intensity increased. This confirms the value of domain-aligned datasets and backbone selection in winter conditions. All annotated data and source code will be published online to support future research.