Computer vision (CV)-based safety monitoring has been widely applied at construction sites. However, this method requires large, diverse, and accurately labeled training data, which is difficult or expensive to collect from real-world environments. To address this concern, this paper introduces a synthetic data generation methodology driven by game engines, facilitating the simulation of diverse construction scenarios from varying distances and viewpoints. Subsequently, a CV model is trained on hybrid datasets encompassing both synthetic and real-world data, thereby evaluating its viability for safeguarding construction workers by particularly detecting small-sized personal protective equipment. Through the incorporation of synthetic data, a detection performance enhancement of up to 30.4%p is achieved, an encouraging outcome with substantial potential to impact worker safety. The outcome of this investigation possess the capacity to refine the safety and overall welfare of construction workers, delivering a cost-efficient and streamlined means to train CV models tailored for safety monitoring purposes.