| 曹伟,王晓勇,刘咸祥.基于改进YOLOX和自适应运动关联的团队体育视频多目标跟踪框架[J].井冈山大学自然版,2024,45(5):81-88 |
| 基于改进YOLOX和自适应运动关联的团队体育视频多目标跟踪框架 |
| A MULTI-OBJECT TRACKING FRAMEWORK FOR TEAM SPORTS VIDEOS BASED ON ENHANCED YOLOX AND ADAPTIVE MOTION CORRELATION |
| 投稿时间：2024-04-16 修订日期：2024-06-19 |
| DOI：10.3969/j.issn.1674-8085.2024.05.011 |
| 中文关键词: 多目标跟踪 YOLOX 自适应运动状态关联 深度学习 团队体育视频 |
| 英文关键词: multi-object tracking YOLOX adaptive motion state association deep learning team sports videos |
| 基金项目:2021年度高等学校省级质量工程项目(2021jxtd259) |
|
| 摘要点击次数: 217 |
| 全文下载次数: 450 |
| 中文摘要: |
| 当前主流多目标跟踪（MOT）方法主要针对行人或车辆等目标的规则运动模式，不适用于团队体育视频中非线性、不规则运动的目标跟踪。针对该问题，提出了基于改进YOLOX和自适应运动状态关联的团队体育多目标跟踪（TS-MOT）深度学习框架。对原YOLOX检测网络进行改进，在主干网络末端集成相似度注意力模块（SimAM），提高特征提取的表征能力，改善颈部组件中特征融合过程。在损失函数中使用变焦损失和广义交并比损失，提高对目标姿态变化和相互重叠的检测能力。此外，在数据关联中，使用具有弱相关性的多个运动状态对目标运动进行表征，通过运动状态相似度计算关联每个目标和轨迹，动态提高轨迹连续性和对意外运动变化的鲁棒性。实验结果表明，所提方法在团队体育MOT任务中表现较好，在SportsMOT数据集上取得了89.4%和78.2%的IDP和IDF1，优于其他先进的MOT方法。 |
| 英文摘要: |
| Currently, multi-object tracking (MOT) methods primarily focus on tracking targets with regular motion patterns, such as pedestrians or vehicles. These methods are not suitable for tracking targets with nonlinear and irregular movements in team sports videos. To address this issue, a deep learning framework named as team-sports multi-object tracking (TS-MOT) based on enhanced YOLOX and adaptive Motion Correlation is proposed. Improvements are made to the original YOLOX detection network by integrating a similarity-based attention module (SimAM) at the end of the backbone network to enhance the feature extraction capability and improve the feature fusion process in the neck component. VariFocal loss and generalized intersection over union (GIOU) loss functions are employed to enhance the detection capability for pose variations and mutual occlusions of multiple targets. Furthermore, in data association, multiple motion states with weak correlations are utilized to characterize the target motions. The adaptive associations of the targets and trajectories based on similarity calculations of motion states dynamically improves trajectory continuity and robustness to unexpected motion changes. Experimental results demonstrate the effectiveness of the proposed method in team sports MOT tasks, achieving 89.4% IDP and 78.2% IDF1 scores on the SportsMOT dataset, outperforming the other advanced MOT methods. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
