基于改进YOLOX-S的足球比赛视频目标检测方法

Improved YOLOX-S-based video target detection method for football matches

为了提升足球赛事水平,催生出足球新战术,识别足球巨星梅西和足球的位置,为进一步的跟踪提供良好的基础,提出了一种基于改进YOLOX-S的足球赛事目标检测方法.使用Pseudo-IoU度量,改进了YOLOX-S中的正样本初步筛选机制,将更标准化和准确的分配规则引入到YOLOX-S无锚检测框架.在损失函数中使用了Focal Loss,以平衡难易样本.实验结果表明,相较于YOLOX-S模型,所提模型具有更好的综合表现,足球类别平均精度为79.8%,梅西类别平均精度为72.6%,平均精度均值为76.2%.

In order to improve the level of football matches,give birth to new football tactics,identify the position of football superstar Lionel Messi and football,and provide a good foundation for further tracking,an improved YOLOX-S target detection method for football matches is proposed.The Pseudo-IoU metric is used to improve the preliminary screening mechanism of positive samples in YOLOX-S,and more standardized and accurate allocation rules are introduced into the YOLOX-S anchor free detection framework.In addition,Focal Loss is used in the loss function to balance the difficulty and easy samples.The experimental results show that the proposed model has better overall performance compared with the YOLOX-S model,with an average accuracy of 79.8%for soccer category,72.6%for Messi category,and 76.2%for the mean accuracy.