Camera-Based Perception under Domain Shifts

Licentiatavhandling, 2025

Autonomous Mobile Robots (AMRs) have emerged as a promising solution for efficient material handling in internal logistics. However, deployment in dynamic, human-shared environments presents major challenges. A core difficulty lies in perception, which must reliably interpret complex scenes to enable safe navigation. This thesis investigates camera-based perception for AMRs, motivated by the availability, low cost, and versatility of cameras. A proposed reference AMR system combining onboard and infrastructure-mounted sensors provides the foundation for exploring key challenges. Two main problems are addressed: the heavy demand for labeled training data in deep learning pipelines, and the difficulty of fusing information from multiple cameras while ensuring robustness to changes in sensor configuration.

To tackle these challenges, the thesis explores unsupervised domain adaptation (UDA) as a unifying strategy. UDA enables models to transfer knowledge from labeled simulated datasets to unlabeled real-world data and to adapt across different multi-camera setups without requiring additional labels. Building on this principle, we propose methods that exploit reliable pseudo-labels and data augmentation to mitigate domain shifts.

The methods are evaluated on benchmarks for semantic segmentation and multi-view pedestrian detection, showing improved performance under domain shifts without extra labeled data. This enables perception systems to rely more on simulated datasets and to adapt more readily to new scenarios, reducing annotation costs. By advancing both monocular and multi-view perception and proposing a reference architecture, this work supports scalable camera-based AMR systems and takes a step toward their widespread deployment.