Autonomous driving promises to transform transportation by enabling automated perception and navigation in complex real-world environments. Vision perception has emerged as a compelling paradigm for autonomous driving, using visual inputs to provide dense cues for semantic understanding and spatial reasoning. However, its reliance on optical inputs renders such systems vulnerable to adverse weather, such as heavy rain and dense fog. These conditions corrupt the sensor observations through visual degradations, disrupting downstream perception and planning models and compromising driving safety.
To address these challenges, this thesis advocates a data-centric framework structured around two primary directions: \textit{data enhancement} for recovering clean inputs from corrupted observations, and \textit{data generation} for creating synthetic driving scenarios for training augmentation. To realize this framework, this thesis delivers a suite of generative and spatial modelling methodologies for both data enhancement and data generation, progressing from 2D image-level processing to 3D scene-level representations. For data enhancement, I first introduce a unified image enhancement framework capable of mitigating diverse visual degradations within a single network architecture. I then extend enhancement from image space to 3D scene reconstruction by explicitly decoupling volumetric atmospheric scattering and camera lens occlusions, enabling the recovery of cleaner scene geometry and appearance under adverse weather. Along the data generation direction, I first develop an LLM-assisted diffusion framework that incorporates autonomous driving domain knowledge into generative modelling and synthesizes corner-case scenarios to improve domain-generalized semantic segmentation. I further advance generation into 3D-aware dynamic scene editing through a controllable weather editing framework operating on a 4D Gaussian field. This framework enables the synthesis of temporally coherent and physically plausible weather scenarios while preserving the underlying 3D scene structure.
Across these settings, the proposed methods achieve competitive or superior performance against relevant task-specific baselines, while the enhanced and generated data improve perception tasks such as semantic segmentation under adverse weather conditions. In conclusion, this thesis advances weather-robust autonomous driving by combining generative priors with spatial scene representations for data enhancement and generation, establishing a scalable data-centric framework for robust visual perception.
