Self-supervised Representation Learning for LiDAR Point Clouds - A Design Science Study of a Self-Supervised Model for Perception in Autonomous Driving

Sammanfattning

Autonomous driving systems are large, complex software systems where accurateenvironmental perception is foundational to safe navigation. Perception systemsoften rely heavily on supervised deep learning models trained on large volumes ofmanually annotated datasets, resulting in performance heavily tied to the qualityof the annotations. Self-supervised learning (SSL) offers a promising alternative byderiving supervisory signals directly from raw, unlabeled data, yet its application toLiDAR point clouds of autonomous driving data remains largely underexplored. Weinvestigate whether a JEPA-based architecture, adapted to operate on LiDAR data,can learn high-quality representations without manual labels, and what that impliesfor the autonomous driving system during its evolution. We examine the meaningfor software quality as well as the impact on the engineering process of buildingand updating the system. Through an iterative design process, we find that ourSSL model substantially outperforms a fully supervised baseline in low label-budgetregimes, and that fine-tuning the pre-trained backbone recovers nearly identicaldetection performance under the full label budget. Our results suggest that SSLpre-training is a viable architectural strategy for reducing annotation dependencyand improving maintainability through backbone reuse, though these benefits comewith meaningful upfront engineering complexity that should be weighed in practice.