3D change detection for autonomous driving

Driven by substantial investments from both industry and research institutions, recent advancements in autonomous driving (AD) have improved the reliability and affordability of these systems. Despite these strides, safety concerns continue to hinder the widespread adoption of AD technology. A fundamental challenge lies in effectively managing the dynamic nature of real-world environments where autonomous vehicles operate. This research addresses a critical consequence of such dynamic environments: discrepancies between the physical world and its pre-existing digital representation, which necessitate adaptive vehicle responses to ensure safety. Change detection algorithms play a pivotal role in bridging this gap. Focusing specifically on 3D change detection, this work seeks to overcome the inherent limitations of commonly used 2D approaches. The proposed pipeline builds upon an unsupervised change detection architecture operating on point clouds. To enable more accurate comparisons, it introduces a novel methodology that generates a dense representation of the static environment from raw sensor data. The pipeline demonstrates superior precision compared to variants based on traditional distance-based change detection algorithms. Finally, this work presents a framework to augment existing autonomous driving datasets for the task of 3D change detection.