Image-Recognition-Based Embedded System for Excavator Bucket Tracking in Construction Sites

Abstract

Construction sites involving excavators often generate substantial amounts of fine dust at the operating position of the excavator bucket. To address the problem, this research proposes a system that tracks the excavator's bucket using only a single camera and an artificial intelligence-based image recognition algorithm, aiming to improve accuracy and efficiency compared to conventional methods that utilize multiple sensors. To enhance the accuracy of image recognition, a bucket dataset containing background images was utilized. Real-time object tracking performance exceeding 30 FPS was achieved by applying a graphics processing unit optimizer. Moreover, a function was implemented to track a specific object when multiple objects with similar characteristics are detected. The system also features a control system that utilizes these functions to apply a pan-tilt motion mechanism to the camera, enabling the tracking of the identified bucket position. Extensive experiments, including image recognition for tracking objects exhibiting various motion trajectories and estimating the position of invisible objects, have been conducted to validate the performance of the system.