Monitoring the vertical progress of reinforced-concrete buildings supports construction management, urban analytics, and seismic exposure classification, yet camera-based floor counting faces two obstacles: public datasets depict almost exclusively completed structures, and the number of structurally finished floors is visually ambiguous while a building is still being erected. We reformulate building-height estimation as discrete floor-count classification from a single street-level facade image and assemble a 29,049-image multi-source corpus centered on the reinforced-concrete urban stock of Kazakhstan, including a 12-month, fixed-viewpoint sequence of 2255 frames that isolates invariance to construction stage, illumination, weather, and season. We formalize a reproducible annotation protocol for three recurring structural ambiguities—incomplete upper floors, rooftop superstructures, and open ground-level pilotis—and propose DINOv2-MSTS, a dual-branch architecture that aggregates multi-scale patch-token statistics from a frozen self-supervised backbone, trained with an Ordinal-Aware Annotation-Uncertainty (OAU) loss for which its Gaussian spread is learned rather than fixed. On the 5359-image Korter + Mendeley 21-category benchmark, the model attains 80% top-1 accuracy, 94% within $\pm 1$ floor accuracy, and 0.28-floor mean absolute error on this saturated 21-category task (a lower bound for buildings of 21 or more floors) using only 1.84 M trainable parameters, 165× fewer than a fully fine-tuned Vision Transformer, which it outperforms by eight accuracy points. On the separate 2255-frame IITU fixed-label robustness probe, it preserves the correct six-floor prediction in 91% of frames (0.09-floor MAE). The corpus, protocol, architecture, and loss together provide a reproducible benchmark for construction-stage building monitoring.

В работе представлена интеллектуальная система мониторинга строительных площадок, разработанная с применением компьютерного зрения и машинного обучения для повышения безопасности и эффективности. Основное внимание уделено автоматическому распознаванию средств индивидуальной защиты (СИЗ) и выявлению нарушений техники безопасности. Использовались нейросетевые модели YOLOv11 и ResNet-50 для детекции и классификации объектов. Облачная обработка данных выполнялась с помощью платформы Lightning AI. При испытаниях на реальных данных си-стема показала точность обнаружения работников и СИЗ на уровне 70–75 % (mAP@0.5), а классификация нарушений достигла 92 %. Однако при более строгой оценке (mAP@0.5:0.95) точность снизилась до 35 %, что указывает на сложности в сложных сценах и при частичном перекрытии объектов. Наибо-лее частыми ошибками стали неверное распознавание касок и защитных очков. Система снижает влияние человеческого фактора и способствует улучшению контроля за соблюдением техники безопасности.

Reliable safety-helmet monitoring in complex industrial environments remains challenging because open-vocabulary detectors are sensitive to prompt wording and changing scene conditions. This study proposes a context-aware framework for YOLO-World and YOLOE that systematically generates color-specific prompts; selects prompts according to the camera angle, time of day, and subject distance; and combines detections using AP-weighted Weighted Box Fusion. The framework was evaluated using strictly separated calibration and test partitions of aerial IITU footage and ground-level CHV construction images. Prompt formulation had the greatest effect on detection performance, with mean AP₅₀ ranging from 0.17 to 0.49 across prompt generation strategies. Weighted fusion produced higher point estimates than the single-best-prompt baseline in all 18 model–dataset–color configurations, with a median improvement of 0.0245 AP₅₀, although most individual gains were not statistically resolved. A supervised YOLOv8-L model remained more accurate in-domain, whereas the open-vocabulary framework could recognize an unseen helmet color without updating detector weights. This flexibility nevertheless requires labeled target-site calibration for prompt ranking and fusion weights. The performance was weakest for long-range overhead views and visually confounded helmet colors. The proposed framework is therefore best suited as a condition-aware monitoring and triage tool, with automation restricted to validated operating regimes and human review retained for uncertain cases.

This study presents a comparative analysis of the object detection models YOLOv8n and YOLOv12n in their Nano configurations for real-time construction site monitoring tasks. Key model characteristics were evaluated, including detection accuracy, recall, inference speed, and computational resource consumption across various server configurations. Particular attention was given to the use of the Slicing Aided Hyper Inference (SAHI) method to improve the detection quality of small objects, as well as the impact of scene complexity on system performance. YOLOv12n achieved significantly higher accuracy and recall than YOLOv8n. In addition, the integration of the SAHI method further enhanced small object detection performance. Parallel stream processing showed that servers equipped with Tesla L4 GPUs offered the best trade-off between inference speed and stability under multi-threaded workloads. The inclusion of data augmentation techniques simulating various weather conditions enhanced model robustness, especially for YOLOv8n, in noisy and complex scenes. The findings underscore the clear advantage of YOLOv12n for applications requiring high recall in challenging environments, while YOLOv8n demonstrated strong efficiency on resource-constrained systems. The proposed approaches can be utilized in the development of intelligent monitoring systems for construction sites, industrial facilities, and logistics hubs, including automated access control and real-time employee activity tracking.