基于深度学习的冰壶检测与轨迹追踪

Deep Learning-based Curling Detection and Trajectory Tracking

冰壶运动轨迹捕捉可以数字化还原比赛过程中冰壶的运动曲线,不仅有利于观众更好了解比赛情况,还可以帮助运动员对当前局势做出判断。运动员与冰壶之间的交互会导致冰壶频繁地被刷子或者运动员身体遮挡,所以从视频序列中直接对移动的冰壶进行追踪是非常困难的。论文针对冰壶比赛场景,通过优化深度学习中的目标检测算法,使用结合冰壶运动特征以及外表特征的多目标追踪方法实现了对冰壶目标的捕捉,坐标的快速校准以及冰壶运动轨迹的合成。论文还提供了一个实际冰壶比赛场景下的标注数据集,弥补了冰壶目标检测与追踪的大型数据集的欠缺。在真实冰壶场景数据集上,与现有的冰壶目标检测与追踪算法相比,实验结果证明,所提方法能够高效完成冰壶追踪任务并且具有很高的准确率。

Curling movement track capture can digitally restore the curling motion curve,which is not only conducive to the au-dience to better understand the competition situation,but also help players to make a judgment on the current situation.Robustly tracking a moving stone from curling video sequences is difficult because the stone is frequently hidden by the brushes held by the players and the players'bodies when players interact with stone.By optimizing the detection algorithm in deep learning,this paper realizes the capture of curling objects,the fast calibration of coordinates and the synthesis of curling trajectories.Multi-target track-ing method combined with curling motion characteristics and curling appearance.This paper also provides an actual curling dataset,which makes up for the lack of large-scale data set of curling.Compared with the existing methods,the experimental results show that the proposed method can efficiently perform tracking task with high accuracy and recall.