AUTHOR=Zhang Pan , Luo Jiangtao 

TITLE=Player detection method based on scale attention and scale equalization algorithm

JOURNAL=Frontiers in Neurorobotics

VOLUME=Volume 17 - 2023

YEAR=2023

URL=https://www.frontiersin.org/journals/neurorobotics/articles/10.3389/fnbot.2023.1289203

DOI=10.3389/fnbot.2023.1289203

ISSN=1662-5218

ABSTRACT=Modern multi-scale object detection methods heavily depend on integrating implicit features across various scales, often interlinked with the dataset's scale statistics. This reliance results in inadequate detection performance for team ball games players with a limited sample volume, particularly leading to missed detections for players with smaller bounding boxes and a reduction in the accuracy of detecting players with larger bounding boxes. Firstly, to tackle the aforementioned challenges, we introduce a multi-scale attention mechanism that minimizes reliance on scale statistics. This mechanism leverages the created SIoU (Similar to Intersection over Union) label to explicitly represent multi-scale features. This guides the training of multi-scale attention network modules, encompassing two levels of granularity, and thereby assisting in reducing the probability of target miss detection. Secondly, a scale equalization algorithm that can be integrated into SIoU labels is proposed to enhance the detection ability of multi-scale targets in imbalanced samples. Comparative experiments were conducted on the dataset containing basketball, volleyball, and ice hockey matches to validate the performance of the proposed method. Its relative optimal method improved the detection accuracy of players labeled with smaller and larger scale bounding boxes by 11%, 7%, 15%, 8%, 9%, and 4%, respectively.