<p>The rapid movement of athletes and complex backgrounds in sports scenes can easily lead to image blur, seriously reducing the accuracy, real-time performance and generalization capabilities of target detection and posture detection, and hindering applications such as event analysis and training evaluation. The research aims to improve the accuracy, real-time performance and generalization ability of sports scene image deblurring and human posture detection. A lightweight deblurring network integrating full-dimensional dynamic convolution is combined with a bidirectional feature enhancement fusion module and a multi-objective constraint loss function to complete network optimization. A global-local attention network is designed to enhance the pose estimation algorithm network. Specifically, the global-local attention neck module and the global-local attention block module are reconstructed as dual core modules, and target tracking, differential data enhancement and unbiased processing strategies are integrated to complete the network settings. The results showed that the peak signal-to-noise ratio of the proposed deblurring algorithm reached 30.45dB, the structural similarity was 0.93, and the texture restoration fidelity reached 0.92. The average posture detection accuracy was close to 90%, the recall was stable at 95%, and the minimum joint point positioning deviation was 3.85 pixels, which was better than that of the mainstream comparison algorithm and could adapt to complex sports scenes. The overall inference delay was controlled within 45.9ms, which can meet the needs of real-time applications. The research adaptively processes various types of motion blurring in motion scenes by integrating full-dimensional dynamic convolution. This effectively solves the problem that traditional fixed convolution is difficult to adapt to multiple types of blurring. Based on the Global-Local Attention Network architecture, it enhances the feature response of key points and suppresses background interference, significantly outperforming mainstream comparison algorithms. It controls the inference delay within 45.9ms, balancing high accuracy and strong real-time performance. It provides precise and efficient technical support for scenes such as sports event analysis and athlete movement assessment, promoting the practical application of computer vision in the sports field, and collaboratively optimizing deblurring and pose detection.</p>