基于DeepSportLab的篮球检测及球员姿态估计研究

Study on Basketball Detection and Player PostureEstimation based on DeepSportLab

为了提升团队体育运动中的智能分析效果,提出通过结合部件强度场和空间嵌入原理来训练多任务,同时实现运动场景中的篮球检测、球员姿态预测和球员实例掩码分割的统一框架(DeepSportLab),以解决团队运动场景的复杂性和特殊性,例如强遮挡和运动模糊。首先,部件强度场提供了篮球和球员的位置信息,以及球员关节的位置。然后,采用空间嵌入技术将球员实例像素与球员各自中心点相关联,并将球员的关节点组合成骨架信息。在DeepSport篮球数据集上进行了验证,并取得了与具有独立任务的单个模型相当的良好性能。

In order to enhance the intelligent analysis effects in team sports,a unified framework named DeepSportLab was put forward,which combines the principles of part strength field and spacial embedding to train multi-task simultaneously.This framework aims to achieve the basketball detection,player pose estimation,and player instance mask segmentation in sports scenes,solving the complexity and particularity of team movement scene such as strong occlusion and motion blur.Within this framework,the part strength field provides positional information for both the basketball and the players,as well as the locations of the player's joints.Then,the spatial embedding technique was employed to associate each player's instance pixels with their respective center points,and the player's joint points were combined to form skeletal information.This method has been validated on the DeepSport basketball dataset and has achieved good performance comparable to individual models with independent tasks.