Real-Time Bucket Pose Estimation Based on Deep Neural Network and Registration Using Onboard 3D Sensor.

Area of Science:

• Robotics and Automation
• Computer Vision
• Geospatial Engineering

Background:

• Bucket pose estimation is crucial for mining excavator intelligence.
• Existing non-visual sensor methods are prone to cumulative errors and short service lives due to vibration and connection issues.
• Occlusion presents a significant challenge for accurate pose estimation.

Purpose of the Study:

• To develop a robust and accurate bucket pose estimation method for mining excavators.
• To address limitations of traditional sensor-based methods, particularly occlusion-induced registration errors.
• To enable real-time and reliable pose estimation for enhanced excavator automation.

Main Methods:

• Optimized a Point Transformer network for precise bucket point cloud semantic segmentation.
• Implemented point cloud preprocessing and continuous frame registration to improve registration accuracy.
• Accelerated the Fast Iterative Closest Point (ICP) algorithm for real-time performance.

Main Results:

• Significantly improved semantic segmentation accuracy for bucket point clouds.
• Reduced registration distance and accelerated ICP algorithm, enabling real-time pose estimation.
• Effectively addressed intermittent pose estimation problems caused by occlusion.
• Validated accuracy and effectiveness through experiments on a custom dataset.

Conclusions:

• The proposed deep neural network and registration-based method enhances bucket pose estimation accuracy and reliability.
• This approach overcomes the limitations of traditional sensors and occlusion challenges in mining environments.
• The method contributes to the advancement of intelligent mining excavators through precise real-time pose estimation.