摘要
以冬奥会的短道速滑比赛场景为例,针对短道速滑中运动员的目标外观差异性小、运动变化快、目标间遮挡频繁等运动特点,设计一个应用于短道速滑场景的多目标跟踪数据集,并提出一种基于轨迹预测增强的多目标跟踪方法.首先计算包围框交并比距离与外观特征余弦距离,联合判断检测响应与跟踪轨迹的相似性解决目标外观相似问题;然后通过跟踪轨迹的全局特征和运动线索恢复被遮挡目标丢失的信息,提高中断轨迹的重关联能力;最后根据检测先验控制新轨迹的初始化,减少噪声检测对轨迹跟踪中身份交换的影响.实验结果表明,与DeepSORT方法相比,所提方法在短道速滑场景中能够稳定地跟踪轨迹,有效地减少了轨迹中断,其中,IDF1提升21个百分点,MOT准确度提高14.3个百分点;该方法在目标差异性小、运动变化快的短道速滑场景中保证长期稳定跟踪,对多目标跟踪在复杂场景中的应用具有启发意义.
Taking the short-track speed skating match of the Winter Olympics as an example,a multiple object tracking dataset applicable to the short-track speed skating scene is designed based on the athletes’motion char-acteristics,such as small distinctions in target appearance,fast motion changes,and frequent mutual occlusion between targets.A multiple object tracking method based on trajectory prediction enhancement was proposed.First,the intersection-over-union of bounding boxes and the cosine distance of appearance features are calcu-lated to jointly judge the similarity between detection and trajectories,solving the problem of similar target ap-pearances.Secondly,the tracking trajectory’s global features and motion clues are used to recover the lost infor-mation of occluded targets,improving the re-association ability of interrupted trajectories.Finally,the detection prior is used to control the initialization of new trajectories,reducing the impact of noisy detection on identity exchange during trajectory tracking.Experimental results show that compared with the DeepSORT method,this method can guarantee stable tracking in the short-track speed skating scene and effectively reduce trajectory interruptions,with an increase of 21 percentage points in IDF1 and 14.3 percentage points in multiple object track-ing accuracy(MOTA).This method ensures long-term stable tracking in short-track speed skating scenes with small differences in targets and fast motion changes.It has inspirational significance for the application of multi-ple object tracking in complex scenes.
作者
刘培刚
王奔
李亚传
崔振东
王君伍
杨少波
李宗民
Liu Peigang;Wang Ben;Li Yachuan;Cui Zhendong;Wang Junwu;Yang Shaobo;and Li Zongmin(School of Computer Science and Technology,China University of Petroleum(East China),Qingdao 266580;Shengli College,China University of Petroleum,Dongying 257061)
出处
《计算机辅助设计与图形学学报》
EI
CSCD
北大核心
2024年第5期786-794,共9页
Journal of Computer-Aided Design & Computer Graphics
基金
国家重点研发计划(2019YFF0301800)
国家自然科学基金(61379106)
山东省自然科学基金(ZR2013FM036,ZR2015FM011).
关键词
深度学习
多目标跟踪
短道速滑
卡尔曼滤波
轨迹预测
deep learning
multiple object tracking
short-track speed skating
Kalman filter
track predict
