Unsupervised Clustering in Football Analysis: A Color-Segmentation and Lighting Adaptation Approach

Abstract

In football match videos, team affiliation is typically identified using unsupervised methods, which distinguish individuals based on unique features. These methods reduce the effort needed for dataset labeling compared to supervised approaches. However, uneven lighting in outdoor football scenes often compromises accuracy. This paper introduces a clustering method leveraging color segmentation combined with illumination equalization to address issues such as large shadows and unknown uniform designs. This method distributes personnel information - distinguishing team A, team B, goalkeepers, and referees - relying solely on color features to achieve precise clustering. Compared to established unsupervised methods, our approach demonstrated superior performance on benchmarks including the Sn-gamestate and Soccernet-Tracing datasets, which contain 81,000 images. Additionally, we developed a shadow correction and color enhancement technique tailored for unevenly lit football scenes. Experimental results show that this method significantly improves clustering accuracy in challenging lighting conditions, boosting the Adjusted Rand Index (ARI) by at least 0.2 and enhancing color restoration markedly.