An Improved YOLOv8 Algorithm for Intensive Small-Target Safety Monitoring in Construction Sites

To address the challenges of detecting intensive small targets in UAV-captured images for construction site safety monitoring, this paper proposes an improved YOLOv8-based algorithm.First, to enhance the extraction of global semantic information and multi-level features, we optimize the FPN sampling by introducing a bidirectional information flow and a formaldehyde feature fusion mechanism.Second, we adopt a twostage training strategy: 1.Freezing the Backbone while training only the detection head in the initial phase. 2. Replacing the original detection head with a Dynamic Head, which incorporates a dynamic fusion mechanism to better adapt to multi-scale features, improving detection robustness for objects of varying sizes.Finally, we enhance the confidence matching mechanism to better distinguish and retain densely packed targets, thereby boosting overall detection performance for intensive small objects.Experimental results on the VisDrone2019 dataset demonstrate that the optimized YOLOv8 effectively meets the demands of small-target detection in complex construction site scenarios.