A Semi-Supervised Domain Adaptation Method for Sim2Real Object Detection in Autonomous Mining Trucks.

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Autonomous trucks are crucial for safety and productivity in open-pit mining.
• High-quality annotated data is essential for training object detection models, but collection is costly and difficult in harsh mining environments.
• Simulation offers a solution for data generation but suffers from Sim2Real domain shift, impacting model performance.

Purpose of the Study:

• To present Adamix, a novel semi-supervised domain adaptation for object detection (SSDA-OD) framework.
• To reduce the Sim2Real domain shift and minimize labeling costs for object detection models.
• To enhance object detection performance in challenging open-pit mining environments.

Main Methods:

• Adamix utilizes a mean teacher architecture with two novel modules: progressive intermediate domain construction (PIDC) and warm-start adaptive pseudo-label (WSAPL).
• PIDC employs a mixup strategy to create intermediate domains, reducing source domain bias and preventing overfitting.
• WSAPL uses adaptive thresholds for pseudo-labeling to mitigate detection errors during training.

Main Results:

• Adamix demonstrated superior domain adaptation performance in a Sim2Real scenario, outperforming state-of-the-art methods.
• The framework achieved higher mean average precision (mAP) compared to existing approaches.
• Active learning integration required 50% less labeled data, significantly reducing real-world data collection needs.

Conclusions:

• Adamix offers a more efficient solution for object detection in open-pit mining by minimizing reliance on expensive real-world data.
• The proposed framework effectively addresses the Sim2Real domain shift challenge.
• Adamix enhances the practicality and cost-effectiveness of deploying autonomous systems in industrial settings.