Embedding Weather Simulation in Auto-Labelling Pipelines Improves Vehicle Detection in Adverse Conditions.

Area of Science:

• Robotics
• Computer Vision
• Machine Learning

Background:

• Deep learning models are crucial for robotic perception but require extensive annotated data.
• Data collection and annotation are significant bottlenecks, particularly for outdoor robots facing diverse conditions.
• Current methods struggle with the vast data needs for real-world robotic deployments.

Purpose of the Study:

• To develop an unsupervised method for training neural networks for vehicle detection in robotic systems.
• To reduce the dependency on human supervision and manual annotation for robot training data.
• To enhance the robustness of robotic perception systems against adverse weather.

Main Methods:

• Utilized a hand-coded algorithm for initial car detection in LiDAR scans under favorable conditions.
• Implemented a tracking method and a weather simulator to generate diverse training data.
• Employed an offline, simulator-in-the-loop approach for automatic data annotation and training.

Main Results:

• The proposed pipeline effectively trains neural networks for vehicle detection without manual annotation.
• The weather simulator significantly improves the robustness of the trained detector.
• The method substantially reduces data collection and annotation efforts for robotic systems.

Conclusions:

• Automated data annotation pipelines are feasible and highly beneficial for deploying deep learning in robotics.
• This approach enables the integration of advanced perception capabilities without extensive manual data preparation.
• The developed framework and provided resources facilitate further research in autonomous systems.