To Örebro University

Robust 2D and 3D perception under extreme conditions, such as adverse weather, dust, or smoke, is essential for autonomous vehicles in unstructured environments like forests, construction sites, and mines. Although Light Detection and Ranging (LiDAR) and cameras provide high-resolution navigation data, rain, snow, and fog introduce noise and signal attenuation that might degrade sensor data. This thesis systematically investigates sensor degradation under adverse weather conditions, quantifies its impact on state-of-the-art perception, develops algorithmic enhancements, and integrates weather awareness into the autonomy pipeline.

This research makes three primary contributions. First, a comprehensive evaluation quantifies the impacts of adverse weather on 2D and 3D object detection, localization, and Simultaneous Localization and Mapping (SLAM). Results show that point cloud filtering benefits LiDAR perception, but standalone image denoising often impairs object detection. Crucially, the study highlights that weather intensity strongly dictates sensor reliability, a factor rarely integrated into current navigation pipelines.

Second, the thesis presents algorithmic improvements to classic Normal Distribution Transform (NDT) registration; two modified variants—Heavily Broadened Likelihood NDT (HBL-NDT) and Overlapping Grid Cells NDT (OGC-NDT)—are introduced to mitigate discretization artifacts and noise, significantly increasing point cloud registration success rates. In unstructured environments, the research demonstrates that leveraging domain knowledge is important. For visual perception, the work establishes that training models on real or synthetically augmented weather data outperforms denoising pipelines and proposes weather augmentation techniques to combat data scarcity.

Third, to address intensity-dependent degradation, the thesis introduces the Video WeAther RecoGnition (VARG) dataset, a video benchmark featuring multi-class weather-type and intensity labels, alongside a context-aware navigation framework that dynamically adjusts filtering, detection thresholds, and control strategies. Collectively, these contributions transform perception systems from static architectures into adaptive, context-aware frameworks, establishing the algorithmic foundations required for safety-critical operations in challenging outdoor environments.