Wearing safety helmets is an effective way to keep construction workers safe on the job. Workers often opt to remove their helmets due to discomfort or a lack of security knowledge, leaving them vulnerable to unsaid threats. A worker's risk of injury from falls, both physical and otherwise, increases when they do not wear a safety helmet. Because detecting whether workers are wearing safety helmets is an important aspect of managing safety on construction sites, a fast and accurate detector is required. However, standard manual monitors aren't particularly popular, and attaching sensors to a safety helmet is a laborious operation. With that in mind, this study presents EM-YOLOv7, a DL method for autonomous safety helmet recognition in real-time that has been developed for use on construction sites. A multi-scale non-local attention module (MSNA) is incorporated into the system to improve detection accuracy in complicated situations. This module enhances feature recovery by aggregating semantic context across multiple scales. Further, to improve detection performance and handle mutual occlusion, a loss function named Wise-IoUv3 (WIoUv3) is used. This method adjusts the loss depending on the overlap between the predicted and ground truth bounding boxes. To train and test the model, we will utilize the safety helmet-wearing (SHWD) dataset, which includes a wide variety of construction workers and whether or not they wore helmets. The goal of this technique is to provide a 99% accurate solution for real-time helmet recognition, which would greatly improve safety management and decrease the likelihood of accidents on construction sites.