High-precision object detection in soccer match videos is of great application significance for intelligent tactical analysis and event broadcasting. However, challenges such as large variations in object scale, motion blur, and dense player occlusion severely limit detection accuracy and robustness. This study proposes a detection algorithm based on improved RT-DETR. A residual Haar discrete wavelet transform (RHDWT) downsampling module is employed to address feature aliasing and loss in tiny objects caused by motion blur. This module utilizes wavelet frequency domain properties to effectively preserve high-frequency edge details. Additionally, the content-aware reassembly of features (CARAFE) module is introduced to replace linear interpolation upsampling. By dynamically predicting reassembly kernels to adjust sampling regions, the semantic alignment accuracy of multi-scale features is improved. Finally, a multi-scale dynamic feature refiner module (MDFRM) is constructed at the input of the decoder. It integrates a convolutional attention mechanism to adaptively enhance feature responses of object regions and suppress background noise interference. Results from experiments conducted on the SoccerNet-Tracking dataset demonstrate that the proposed algorithm outperforms the baseline model RT-DETR and mainstream YOLO series models, yielding the best performance across all accuracy metrics. The precision and recall reach 89.9% and 85.4% respectively, and mAP@50 reaches 84.9%. This method effectively solves the problems of missed detection and false detection in complex soccer match scenes, providing reliable support for intelligent analysis of soccer matches.