High-definition (HD) maps provide centimeter-level spatial accuracy essential for autonomous driving, yet traditional production workflows present significant scalability limitations due to resource-intensive data collection and manual processing. This thesis addresses the critical research gap by investigating the integration of two independently evolving domains: learning-based end-to-end HD mapping and HD map-based localization systems.
Current approaches develop along parallel trajectories, limiting system adaptability as real-world autonomous systems require both continuously updated maps and precise localization within evolving environments. Through comprehensive literature survey, experimental validation, and architectural framework design, this research demonstrates the fundamental interdependence between mapping and localization performance.
The comparative analysis reveals that both HD and standard-definition (SD) map-based approaches converge on transformer-based Bird's Eye View representations and cross-modal attention mechanisms, indicating that semantic-geometric correspondence learning represents the core technical challenge. HD map-based methods achieve centimeter-level accuracy under constrained conditions while SD map-based methods provide meter-level precision with superior tolerance to large pose uncertainties.
Experimental validation establishes the fundamental interdependence between mapping and localization through systematic pose perturbation analysis that simulates real-world GPS/IMU positioning uncertainties. Localization experiments demonstrate that training with realistic pose uncertainties significantly enhances performance stability compared to idealized scenarios, while HD mapping experiments reveal a dual-threshold degradation pattern where fine-scale perturbations induce gradual performance reduction and large-scale perturbations precipitate catastrophic collapse. These findings establish that accurate pose estimation enables superior mapping through effective historical map integration, whereas pose uncertainties fundamentally compromise spatial correspondence and prior information retrieval.
The proposed end-to-end architecture enables simultaneous HD mapping and localization through joint optimization, featuring hybrid prior map integration, dual-layer storage balancing stability with environmental responsiveness, and decoupled pose estimation maintaining computational efficiency.
This research provides the first comprehensive survey bridging HD mapping and localization domains, establishes quantitative performance benchmarks, and identifies optimal integration strategies. The findings inform next-generation autonomous driving systems capable of maintaining accuracy while adapting to dynamic operational environments.