Incheon National University Researchers Find Solution for Reliable Excavator Tracking in Real-World Construction Environments

Researchers develop an automated, reliable-based multi-camera strategy applicable to simultaneous occlusion scenarios in dynamic construction environments

INCHEON, South Korea, May 4, 2026 /PRNewswire/ -- A recent study published in Automation in Construction by researchers from Incheon National University exploits a novel approach to improving excavator tracking performance under real-world conditions. By integrating deep learning-based instance segmentation with an automated, reliability-based multi-camera strategy, this study addresses one of the most persistent challenges in construction monitoring—frequent occlusions caused by dynamic site activities. In addition, the researchers propose a frame-level reliability estimation process that automatically identifies unreliable tracking results.

Despite significant advances in vision-based equipment tracking, frequent occlusions caused by multiple interacting machines continue to degrade tracking accuracy on construction sites. While previous studies have explored multi-camera approaches, they often assume that at least one camera maintains a clear view at all times. In practice, however, such conditions are rarely guaranteed. Even with multiple CCTV systems installed, simultaneous occlusions across cameras frequently occur, making it difficult to identify which camera provides the most reliable view at any given moment.

To fill this knowledge gap, a research team led by Professor Choongwan Koo from the Division of Architecture & Urban Design at Incheon National University, Republic of Korea, has recently proposed an automated, reliability-based multi-camera strategy for excavator tracking. This study was made available online on 9 October 2025, and have been published in Volume 181 of Automation in Construction on 1 January 2026.

According to Prof. Koo, "The primary practical contribution of this study is the development of a quantitative and automated framework for selecting the most reliable camera under simultaneous occlusion conditions in real-world construction environments."

The proposed approach introduces a quantitative reliability framework based on occlusion regions (e.g., arm and body) and occlusion ratios. Unlike conventional methods that primarily focus on improving detection or tracking algorithms, this strategy enables direct assessment of tracking reliability at the frame level. Notably, this study identifies critical occlusion thresholds—0.7 for the excavator arm and 0.5 for the body—beyond which tracking performance significantly deteriorates. These thresholds offer clear and actionable criteria for practitioners to assess the quality of vision-based monitoring data.

Beyond methodological contributions, the proposed strategy also offers significant practical and economic benefits. By improving the continuity and accuracy of tracking results, the approach enhances the reliability of equipment operation logs, which are essential for carbon emission estimation and regulatory reporting. This, in turn, can help reduce administrative burdens and financial risks, such as penalties or revalidation costs. Furthermore, by leveraging existing camera infrastructure and evaluating reliability in real time, the proposed approach minimizes the need for additional camera installations, thereby reducing capital expenditure and avoiding complications related to power supply, networking, and site constraints. The automated camera selection process also reduces reliance on manual intervention, leading to lower labor and operational costs.

"Overall, our work provides a practical and scalable solution to improve the reliability and field applicability of vision-based construction monitoring systems, with potential extensions to productivity analysis, activity recognition, and carbon emission monitoring," concludes Prof. Koo.

Reference
Title of original paper: Automated reliability-based multi-camera strategy for excavator tracking under dynamic occlusion using deep learning with instance segmentation
Journal: Automation in Construction
DOI: 10.1016/j.autcon.2025.106589

About Incheon National University
Website: https://www.inu.ac.kr/sites/inuengl/index.do?epTicket=LOG

Media Contact:
Jong Jin Lee
82-32 - 835 - 9325
[email protected]

SOURCE Incheon National University