In modern petroleum engineering, ensuring operational safety at drilling sites is of critical importance. Visual object detection plays a key role in intelligent safety monitoring systems by enabling real-time supervision of personnel and equipment. However, safety-critical targets in drilling scenes are often small, partially occluded, and embedded in cluttered environments, leading to decreased detection accuracy and potential safety risks. Existing convolutional neural networks (CNN)-based detectors, although effective in natural scenes, often exhibit limited robustness under such complex industrial conditions. To address these challenges, this paper proposes MSA-DETR, a Transformer-based detection framework designed to enhance multi-scale perception in drilling monitoring scenarios. By improving the ability to capture both global contextual information and fine-grained visual cues, the proposed approach enhances sensitivity to safety-relevant objects. Extensive experiments conducted on two real-world drilling monitoring datasets demonstrate that MSA-DETR consistently outperforms state-of-the-art detection methods, providing more reliable visual perception for petroleum safety management and accident prevention.