The safety of construction workers is becoming an increasingly important issue for various building sectors. Workers' compensation claims may be reduced if more of them wore safety helmets, yet improper use of helmets is a common problem on construction sites. Therefore, a system that can automatically identify safety helmets using computer vision is crucial. Although many studies have focused on helmet detection in general, few have specifically addressed its use in construction locations. In this paper, we have designed an onsite automatic supervision model for helmet detection using an improved convolutional neural network to determine whether people are wearing helmets or not on the construction site. In the proposed model, firstly, pre-processing of the input data image is done to enhance and extract the appropriate features from it. To achieve this goal, histogram equalization and the local binary pattern (LBP) algorithm are deployed. After that, a convolutional neural network (CNN) is applied for helmet detection, which classifies whether the helmet is worn or not. The proposed model is simulated on the standard dataset. The dataset contains two classes. The first class contains images without helmets, and the second class contains helmet images. Further, the visual quality of the input image and images generated after enhancement and feature extraction is shown in the qualitative analysis, whereas accuracy, precision, recall, and F-score are measured in the quantitative analysis. The result shows that the proposed model outperforms both classes and achieves high accuracy (0.95652).