Optimizing Sprint Start Performance: A Study on Human Pose Estimation with Frequency Domain Integration

To address the issue of sprinting beginners making errors in their starting technique due to the lack of adequate guidance, this research proposes an improved HRNet-based algorithm, known as FD-HRNet. The model introduces a frequency domain branch and replaces traditional residual blocks with a Fast Fourier Residual Block with a Reweighted Complex Matrix (FFT ResBlock with Wc), hence enhancing its ability to extract frequency domain features. In addition, a specialized dataset was developed focusing on the technical characteristics of sprint starts, accompanied by a set of evaluation criteria to assess the accuracy of sprint start actions. The experimental results revealed that the implementation of FD-HRNet enhanced the recognition accuracy by 1.7% on the COCO dataset and exhibited superior practical performance on the sprint start dataset. This improvement provides a more accurate identification of beginners’ errors, hence offering valuable guidance for their correction.