Bypassing the Labeling Bottleneck : Automated Ground Truth Generation and Parameter Optimization for LiDAR Perception Software

Sammanfattning

As Light Detection and Ranging (LiDAR) systems become increasingly important in modern intelligent traffic management, deploying these systems’ underlying perception software requires rapid, precise parameter calibration to ensure accurate object detection in new operational environments. Traditionally, this configuration process relies on manual trial-and-error, a paradigm that is inherently time-consuming, subjective, and computationally irreproducible. Furthermore, optimizing LiDAR systems requires high-fidelity reference datasets, creating a critical labeling bottleneck where human manual annotation becomes a constraint on system scalability. This study adopts a Design Science Research approach to develop, implement, and comparatively evaluate two semi-automated data annotation workflows alongside two automated parameter tuning frameworks for Flasheye’s operational ecosystem. To alleviate the data labeling constraint, a machine learning-driven Active Learning (AL) pipeline utilizing an Extra Trees Regressor committee variance metric was evaluated. This was evaluated against a deterministic, weak-supervision strategy based on temporal piecewise calibration. The resulting high-fidelity ground truths served as the basis for black-box parameter optimization, where an evolutionary Genetic Algorithm (GA) was pitted against a probabilistic Bayesian Optimization (BO) heuristic. The empirical findings indicate that while the Active Learning approach offered exceptional modeling flexibility, the most effective architecture was the Programmatic Labeling pipeline paired with TPE Bayesian Optimization. This combined framework required down to one-sixth of the operational effort related to data labeling while achieving up to a 69% increase in tracking performance compared to the human-engineered baseline. Ultimately, this thesis contributes a scalable, automated framework that systematically combines data processing with algorithmic parameter tuning, providing a foundation for autonomous perception-device calibration in infrastructure-based surveillance networks.