In construction areas, where workers frequently come into close proximity with heavy machinery, worker safety is of utmost importance. The present research enhances the YOLOv8 architecture to present a more sophisticated object detection method. The proposed enhanced methodolgy combines Convolutional Block Attention Modules (CBAM) and Squeeze-and-Excitation (SE) mechanisms for enhanced spatial and channel-wise attention, in addition to Scylla IoU (SIoU) loss for enhanced bounding box regression. In order to assess spatial safety, the system also integrates the ZoeDepth model for depth estimation and worker-machinery proximity assessment. The evaluation is conducted using a custom dataset called CSOD-24, which consists of 10,000 annotated frames taken from real Maharashtra construction sites. There are four object classes associated with safety compliance in this dataset. The enhanced model maintains real-time inference capabilities while consistently outperforming the baseline YOLOv8 in detecting small and partially occluded objects. Its suitability for intelligent, automated construction site safety monitoring is confirmed by the experimental findings.