Failure to use proper personal protective equipment can lead to fatal consequences or complex accidents. Binary helmet detection is a very active research topic that involves detecting whether a worker is wearing a safety helmet or not. Despite the great interest, almost all of the work focuses on the analysis of images acquired in construction environments. To support mining safety monitoring, we propose a new dataset collected from mining environments, consisting of 1,510 images labeled for helmet and head object detection. Furthermore, we propose an uncertainty-driven model aiming to improve the performance by reducing the false positive rate. The effectiveness of the proposed method was demonstrated on the proposed dataset, and also by a performance analysis between cross-datasets strategy. The proposed method provides 95.3% in terms of mAP and 91.4% in terms of F1, which is a promising performance for binary helmet detection in real solutions applied to mining environments. This work contributes to better control of the use of safety helmets to reduce the complexity of head injuries related to work-related accidents, thus protecting workers and guaranteeing the continuity of the operational process.