Construction of Stretching-Bending Sequential Pattern to Recognize Work Cycles for Earthmoving Excavator from Long...

Area of Science:

• * Construction Engineering
• * Computer Vision
• * Machine Learning

Background:

• * Accurate counting of earthmoving excavator work cycles is crucial for project productivity assessment.
• * Existing computer vision methods struggle with long video sequences and visually similar atomic actions.
• * Previous approaches often overlook real-world factors like driver misoperation.

Purpose of the Study:

• * To develop an effective method for recognizing earthmoving excavator work cycles in long video sequences.
• * To enhance the accuracy of excavator productivity calculations by addressing limitations in current computer vision techniques.
• * To incorporate real-world factors, such as driver misoperation, into work cycle recognition.

Main Methods:

• * Development of a Stretching-Bending Sequential Pattern (SBSP) by combining visually similar atomic actions.
• * Utilizing a deep learning-based Single-Shot Detector (SSD) to recognize "Stretching" and "Bending" atomic actions.
• * Employing Intersection over Union (IOU) for atomic action association and a time-based filter to handle driver misoperation.

Main Results:

• * The proposed SBSP method effectively recognizes earthmoving excavator work cycles in real-time within long video sequences.
• * The method demonstrates the ability to accurately calculate excavator productivity.
• * The time-based filtering successfully mitigates issues caused by abnormal work cycles due to driver misoperation.

Conclusions:

• * The SBSP method offers a robust solution for automated work cycle recognition of earthmoving excavators.
• * This approach significantly improves the accuracy and reliability of productivity calculations in earthmoving projects.
• * The integration of real-world operational factors enhances the practical applicability of the system.