This paper proposes integrating a Partial Convolution (PConv) module into the backbone network of YOLOv8 to enhance robustness in detecting side-view athletes on running tracks. Addressing common challenges in track scenarios—partial occlusion, lateral poses, motion blur, and background interference—we designed a direction/occlusion-aware PConv variant. This variant masks and normalises obscured or featureless regions during feature extraction, thereby improving the detector's response to lateral human silhouettes. This paper presents the mathematical definition of PConv, its integration with YOLOv8, training strategies, and experimental design. An ablation study analyses the impact of inserting PConv at different positions (backbone, neck, FPN). Finally, we discuss its advantages, limitations, and future research directions.