Deep Learning-Enabled Construction Progress Monitoring and Schedule Deviation Analysis Using BIM and Computer Vision

Delayed identification of schedule deviations is a persistent driver of cost overruns and productivity loss in construction, particularly in small and medium residential projects where progress tracking relies heavily on manual inspection and engineer judgement. Building Information Modelling (BIM) improves planned-progress visualisation through four-dimensional (4D) scheduling; however, capturing actual site execution in a structured and comparable form remains challenging. This study developed and validated a lightweight, photograph-based framework for extracting element-level construction progress from ordinary two-dimensional site photographs and comparing it with a BIM-derived planned schedule baseline without full as-built geometric reconstruction or specialised sensing hardware.

The implemented framework combines a BIM-based planned baseline with YOLOv8n-seg instance segmentation and a formwork lifecycle-to-completion mapping. The YOLOv8n-seg model detected reinforced-concrete columns, beams, and slabs, achieving an overall box mAP@0.5 of 0.849 and mask mAP@0.5 of 0.862; the Column class, which underlies the schedule-deviation validation, achieved 0.902 for both metrics. A supplementary Formwork-class detector supported the lifecycle mapping. Observable formwork states were mapped to completion levels of 25% for shuttering, 75% for concreting, and 100% after deshuttering. Detected element counts and completion percentages were compared element-wise with the dated BIM-derived planned baseline to calculate element-count deviation (ΔE) and completion-percentage deviation (Δ%), classifying status as Ahead, On Track, or Delayed.

The framework was validated on first-floor column construction of the Naser Apartment, a G+2 residential building in Malappuram, Kerala, using three site observations in February 2026. On 25 February, sixteen of twenty-one planned formwork elements were detected, producing ΔE = −5 and Δ% = −23.8% and correctly classified as Delayed. On 26 February, all twenty-one planned elements were detected, indicating recovery to On Track, corroborated on 27 February following the transition from Formwork to Column detection after deshuttering. The results demonstrate the feasibility of using ordinary site photographs and deep-learning-based detection for element-level planned–actual schedule comparison under genuine field conditions. Given the single-project validation scope, the findings indicate practical potential rather than statistical generalisability, and broader validation is required.