Rock Particle Motion Information Detection Based on Video Instance Segmentation.

Area of Science:

• Geotechnical Engineering
• Computer Vision
• Particle Mechanics

Background:

• Accurate detection of rock particle motion is essential for understanding particle behavior, guiding construction, and preventing geological hazards.
• Existing methods may struggle with precise motion analysis, especially under dynamic conditions like vibration loads.

Purpose of the Study:

• To develop and validate a machine vision method for detecting rock particle motion information under vibration loads.
• To enhance object detection and segmentation accuracy for similar-looking particles and backgrounds.
• To enable precise tracking and quantitative analysis of particle translation and rotation.

Main Methods:

• An improved Mask R-CNN model incorporating an ArcFace loss function for enhanced classification of similar objects and backgrounds.
• Integration of the enhanced Mask R-CNN with Deep Simple Online and Real-time Tracking (Deep SORT) for robust particle detection, segmentation, and tracking.
• Utilizing equivalent ellipse characterization and proportional calibration for quantitative measurement of particle translation and rotation.

Main Results:

• The enhanced Mask R-CNN achieved 93.36% accuracy on a custom dataset and demonstrated advantages on public datasets.
• The combined Mask R-CNN and Deep SORT system effectively performed video instance segmentation with a low ID switching rate.
• Average detection errors for translation and rotation were 5.10% and 14.49%, respectively.

Conclusions:

• The proposed intelligent scheme provides an effective solution for detecting rock particle movement information under vibration.
• The method offers improved accuracy and robustness for particle motion analysis in geotechnical applications.
• This approach has significant potential for applications in engineering construction, geological disaster prevention, and numerical model verification.