An Integrated Computer Vision Framework for Construction Equipment Monitoring, Utilization Analytics, and Decision Support

Effective construction-equipment monitoring requires persistent equipment-level information linking visual observations to utilization and planning decisions. This study develops and evaluates an identity-centered framework integrating ten-class YOLOv8m detection, ByteTrack tracking, AprilTag-assisted Global ID reconciliation, scale-normalized motion-state estimation, utilization accounting, and rule-based reallocation support. The selected detector achieved 90.80% micro-F1 on the held-out ConstructionXC7C benchmark (757 images, eight annotated categories). In a matched comparison, selected-checkpoint micro-F1 reached 92.52%, exceeding the original checkpoint by 16.11 percentage points. Across 6000 experimental video frames, 5757 of 5992 annotated excavator observations were matched, with 99.81% adjacent camera–local ID consistency. A controlled 192-pair simulated-transition study associated all 144 same-object pairs using ideal deterministic identity tokens, compared with 99 using a tuned appearance-only baseline; regression tests verified preservation of accumulated equipment histories. In a separate matched analysis of 1040 observations, temporal activity modeling improved idle-state discrimination and reduced video-level sampled-utilization mean absolute error by 41.92 percentage points, achieving an error of 15.11 percentage points. Controlled verification further demonstrated state-service functionality and reviewable zone-reallocation advisories. These findings provide quantitative evidence of component-level performance and traceable monitoring-to-decision functionality, establishing a technical foundation for subsequent construction-site evaluation.