Generation of Training Data from HD Maps in the Lanelet2 Framework (2407.17409v1)

Abstract: Using HD maps directly as training data for machine learning tasks has seen a massive surge in popularity and shown promising results, e.g. in the field of map perception. Despite that, a standardized HD map framework supporting all parts of map-based automated driving and training label generation from map data does not exist. Furthermore, feeding map perception models with map data as part of the input during real-time inference is not addressed by the research community. In order to fill this gap, we presentlanelet2_ml_converter, an integrated extension to the HD map framework Lanelet2, widely used in automated driving systems by academia and industry. With this addition Lanelet2 unifies map based automated driving, machine learning inference and training, all from a single source of map data and format. Requirements for a unified framework are analyzed and the implementation of these requirements is described. The usability of labels in state of the art machine learning is demonstrated with application examples from the field of map perception. The source code is available embedded in the Lanelet2 framework under https://github.com/fzi-forschungszentrum-informatik/Lanelet2/tree/feature_ml_converter

• The paper details lanelet2_ml_converter’s creation of 'compound labels' to eliminate annotation artifacts and standardize map-derived training data.
• It achieves real-time performance with an average of 3 ms per frame on a standard Intel i7, ensuring prompt data processing for automated driving.
• The framework guarantees traceability by linking each generated label to unique map element IDs, thus enhancing quality control and debugging.

Generation of Training Data from HD Maps in the Lanelet2 Framework

The paper entitled "Generation of Training Data from HD Maps in the Lanelet2 Framework" addresses a critical need in the domain of automated driving: a standardized framework that integrates high-definition maps into machine learning workflows. This work builds upon the Lanelet2 framework, a popular HD map format utilized in both academic and industrial automated driving systems, to support the generation of training labels for map perception tasks. The proposed extension, named lanelet2_ml_converter, focuses on unifying map-based automated driving, machine learning inference, and training, all from a single source of map data and format.

Key Contributions

1. Motivation and Framework Requirements: The paper emphasizes the necessity for a unified HD map framework that is versatile enough to serve both automated driving and deep learning training tasks. It outlines specific requirements such as the generation of training labels from maps, traceability of labels, map validation, and independence of labels from map annotation artifacts.
2. Extension Design and Implementation: A significant portion of the paper details the design and implementation of the lanelet2_ml_converter. The authors introduce "compound labels" to address the issue of map annotation artifacts and ensure consistent, artifact-independent label generation. The implementation achieves real-time performance, critical for online inference and fusion tasks.
3. Traceability and Map Validation: The extension ensures full traceability of generated labels to their originating map elements, leveraging unique IDs for quality control and debugging. In addition, the Lanelet2 validation module is integrated to maintain data consistency and error-free label generation.
4. Support for Multiple Applications: The paper demonstrates how the extension can be utilized to generate training data for various map perception problems. Examples include the generation of vectorized map elements for online HD map construction and topology inference, as seen in models like MapTR and dataset formats like OpenLaneV2.

Technical Details and Numerical Results

The marginal yet critical ability to generate local instance labels in real-time is quantified by the authors, reporting an average single-threaded performance of 3 ms on a standard Intel i7 processor. This performance metric is crucial for practical deployment in dynamic, time-sensitive environments like automated driving systems. The compound label generation technique effectively handles the complexity of merging geometric primitives and ensures robust labeling that scales with dynamic changes in the map data.

Implications and Future Developments

The practical implications of this work are substantial, particularly in the context of automated driving systems that rely on both sensor perception and HD maps. By integrating label generation within the Lanelet2 framework, the lanelet2_ml_converter enhances the ability to train and infer machine learning models in real-time, thus facilitating more adaptive and resilient driving agents.

Theoretically, the paper advances the discourse on HD maps by presenting a unified, extensible solution that bridges the gap between static map data and dynamic machine learning requirements. This holds potential for further research and development, particularly in improving the abstractions and representations of non-road features like traffic signs and signals.

Future work could extend the framework to better support a diverse range of map elements beyond road features, incorporating traffic control elements and enhancing the robustness of path planning and behavior prediction models. Additionally, exploring the integration of this framework with more sophisticated mapping and localization techniques could further elevate its utility and scope.

Conclusion

The "Generation of Training Data from HD Maps in the Lanelet2 Framework" paper substantially contributes to the field of automated driving by presenting a comprehensive extension to the Lanelet2 framework for generating training data. This work addresses existing limitations in HD map frameworks and lays the groundwork for both improved map-based driving and advanced map perception tasks. Through robust implementation and promising application examples, this research sets a foundation for future advancements in the fusion of HD maps and machine learning methodologies within automated driving systems.