Abstract: Object detection technology, as a core link in intelligent perception, holds significant application value in fields such as autonomous driving and intelligent security. To address the limited feature representation capability of traditional DETR models in complex scenes, this study proposes Counting-DETR, an algorithm that integrates object counting information to achieve dual improvements in detection accuracy and inference efficiency. This method adopted a set-based global loss that enforced unique predictions through bipartite matching, utilizing object counting information to assist visual feature enhancement. The counting information was injected into the visual model as global clues to strengthen the reasoning capability of visual features. A Transformer encoder based on global counting information was proposed to extract more comprehensive visual feature representations, guiding the model to obtain more accurate detection results. Experimental results demonstrate that integrating object counting information into the visual model effectively enhances semantic understanding of complex scenes, provides new insights for addressing target missed detection and false detection in dynamic environments, and lays a theoretical foundation for the engineering application of multimodal perception algorithms.