In complex sports scenarios, there are some problems in the recognition and detection of sports targets, such as low detection efficiency and accuracy, and many environmental noises. Therefore, a mechanism of multi attention fusion of YOLOv9 was proposed. Firstly, the multi-scale dilated attention (MSDA) mechanism was introduced at the output end to understand the image at different scales, so as to enhance the ability to obtain features and capture details. Secondly, the deformable large kernel convolution(DLKA) mechanism was introduced at the output end to make the model adapt to the patterns of different data and improve the ability of image recognition to adapt to complex environments. Experiments show that mAP50 and mAP75 increase 1.4% and 1.7% respectively compared with the basic model in athlete detection. It is verified that the algorithm has significant motion target recognition ability, which provides an important foundation for the detection of motion targets in smart construction and smart sports.