| Title |
Tracking Reliability Estimation Model for Construction Equipment under Occlusion - Unreal Engine Simulation and Deep Learning Approach - |
| Authors |
Seongkyun Ahn ; Seungwon Seo ; Choongwan Koo |
| DOI |
https://dx.doi.org/10.6106/KJCEM.2026.27.1.060 |
| Keywords |
Excavator; Occlusion; Tracking Performance (MOTA); Vision-based Tracking Reliability; Unreal Engine; 3D Simulation |
| Abstract |
This study aims to enhance the field applicability of deep learning-based heavy equipment tracking models under occlusion conditions commonly encountered on construction sites. These environments involve dynamic interactions between numerous machines and workers, often resulting in frequent occlusions that cause tracking errors and degrade the reliability of vision-based monitoring data. To address this challenge, the study developed various synthetic scenarios using Unreal Engine, simulating different occlusion scenarios (arm and body) and occlusion ratios. A deep learning-based tracking model was applied to conduct a quantitative sensitivity analysis of tracking performance under these conditions. The analysis revealed that occlusion of the arm led to an average MOTA performance drop of 42.22%, while body occlusion caused a drop of 50.62%. Furthermore, the critical occlusion ratio thresholds?beyond which performance deteriorates sharply?were found to be 0.7 for the arm(α) and 0.5 for the body(β). Based on these findings, this study proposed a frame-level tracking reliability estimation process that quantitatively assesses tracking confidence under varying occlusion conditions. This process automatically flags low-reliability frames, thereby improving data verification efficiency and enabling robust tracking continuity even in partial occlusion environments. The proposed model offers a practical foundation for improving the field applicability of vision-based monitoring systems on construction sites and holds high potential for integration with motion recognition, carbon emission monitoring, and productivity assessment systems in future applications. |