基于改进的Deep-EIoU对多名足球运动员跟踪模型

doi:10.11996/JG.j.2095-302X.2026051001

摘要/Abstract

摘要：

针对多名足球运动员跟踪场景中因外观相似与高速运动引发的身份混合与轨迹中断问题，提出一种基于改进型Deep-EIoU的多目标跟踪(MOT)模型。以Deep-EIoU的检测-匹配框架为基础，构建结合外观一致性与运动连贯性的轨迹级全局优化机制，在初始关联结果上进一步进行轨迹结构修正。首先，通过轨迹分割模块识别并剥离混合身份片段，以消除错误身份传播。其次，通过轨迹重连模块在时空约束下实现断裂轨迹的全局融合，以恢复长时序目标的连续跟踪，从而提升复杂动态场景下的身份保持与轨迹连续性。实验结果表明，改进Deep-EIoU在HOTA，IDF1，AssA，MOTA和DetA等指标上表现优秀，在SoccerNet-Tracking上分别达到89.1%，82.1%，79.4%，96.7%和99.1%，在SportsMOT上分别达到78.1%，80.1%，68.4%，97.1%和88.6%，在身份保持与轨迹连续性方面均优于主流多目标跟踪方法。该方法在足球及其他复杂体育场景中具有较高跟踪准确度，为运动表现分析和战术研究提供了初步技术支撑。

关键词: 足球, 运动员, 多目标跟踪, 深度学习, 外观关联

Abstract:

To address the challenges of identity confusion and trajectory fragmentation caused by similar appearances and rapid motion in multi-football-player tracking scenarios, a Multi-Object Tracking (MOT) model based on an improved Deep-EIoU framework was proposed. Building on the original detection-matching pipeline of Deep-EIoU, the proposed method introduced a trajectory-level global optimization mechanism combining appearance consistency and motion coherence, and the initial association results at the tracklet level were further refined. First, a tracklet splitter module was used to identify and separates mixed-identity segments, eliminating the propagation of erroneous identities. Second, a tracklet connector module performed global fusion of fragmented trajectories under spatio-temporal constraints to recover long-term continuous tracking of targets, thereby improving identity preservation and trajectory continuity in complex dynamic scenarios. Experimental results showed that the improved Deep-EIoU achieved superior performance on multiple metrics, with HOTA, IDF1, AssA, MOTA, and DetA reaching 89.1%, 82.1%, 79.4%, 96.7%, and 99.1% on the SoccerNet-Tracking dataset, and 78.1%, 80.1%, 68.4%, 97.1%, and 88.6% on the SportsMOT dataset, respectively, outperforming mainstream multi-object tracking methods in both identity stability and continuity. The proposed model demonstrated high tracking accuracy in football and other complex sports scenarios, providing a technical foundation for intelligent sports analysis and tactical research.

Key words: football, players, multi-object tracking, deep learning, appearance association

中图分类号:

吴雪, 刘子平. 基于改进的Deep-EIoU对多名足球运动员跟踪模型[J]. 图学学报, 2026, 47(5): 1001-1014.

WU Xue, LIU Ziping. A multi-football-player tracking model based on improved Deep-EIoU[J]. Journal of Graphics, 2026, 47(5): 1001-1014.

图/表 13

参考文献 25

[1]	李永, 林坊, 陈煜昂, 等. 基于深度学习的视觉多目标跟踪综述[J]. 科学技术与工程, 2025, 25(22): 9211-9223.
	LI Y, LIN F, CHEN Y A, et al. A comprehensive review of deep learning-based visual multi-object tracking[J]. Science Technology and Engineering, 2025, 25(22): 9211-9223 (in Chinese).
[2]	李天成, 谢昱昕, 李固冲, 等. 面向多目标跟踪的数据关联方法研究综述[J]. 雷达科学与技术, 2025, 23(1): 10-31.
	LI T C, XIE Y X, LI G C, et al. A survey on data association approaches to multi-target tracking[J]. Radar Science and Technology, 2025, 25(22): 10-31 (in Chinese).
[3]	郭凡, 卢铉宇, 李嘉怡, 等. 基于深度神经网络的目标跟踪算法综述[J]. 航空兵器, 2024, 31(1): 1-12.
	GUO F, LU X Y, LI J Y, et al. A review of target tracking algorithm based on deep neural network[J]. Aero Weaponry, 2024, 31(1): 1-12 (in Chinese).
[4]	张红颖, 贺鹏艺, 彭晓雯. 基于改进高分辨率神经网络的多目标行人跟踪[J]. 光学精密工程, 2023, 31(6): 860-871.
	ZHANG H Y, HE P Y, PENG X W. Multi-object pedestrian tracking method based on improved high resolution neural network[J]. Optics and Precision Engineering, 2023, 31(6): 860-871 (in Chinese). DOI URL
[5]	田嘉意, 李辉, 李赛宇, 等. 基于多重信息融合与轨迹关联修正的多目标跟踪方法[J]. 控制与决策, 2024, 39(6): 1888-1896.
	TIAN J Y, LI H, LI S Y, et al. A multi-object tracking method based on multi-information fusion and trajectory correlation correction[J]. Control and Decision, 2024, 39(6): 1888-1896 (in Chinese).
[6]	霍旭, 盖绍彦, 洪濡, 等. 基于多阶段关联的多目标跟踪算法[J]. 仪器仪表学报, 2023, 44(11): 205-214.
	HUO X, GAI S Y, HONG R, et al. Multi-object tracking algorithm based on multi-stage association[J]. Chinese Journal of Scientific Instrument, 2023, 44(11): 205-214 (in Chinese).
[7]	MILAN A, LEAL-TAIXE L, REID I, et al. MOT16:a benchmark for multi-object tracking[EB/OL]. (2016-03-02) [2025-09-30]. https://arxiv.org/abs/1603.00831.
[8]	DENDORFER P, REZATOFIGHI H, MILAN A, et al. MOT20:a benchmark for multi object tracking in crowded scenes[EB/OL]. (2020-03-19) [2025-09-30]. https://arxiv.org/abs/2003.09003.
[9]	CIOPPA A, GIANCOLA S, DELIEGE A, et al. SoccerNET-tracking: multiple object tracking dataset and benchmark in soccer videos[C]// 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. New York: IEEE Press, 2022: 3490-3501.
[10]	CUI Y T, ZENG C K, ZHAO X Y, et al. SportsMOT: a large multi-object tracking dataset in multiple sports scenes[C]// 2023 IEEE/CVF International Conference on Computer Vision. New York: IEEE Press, 2023: 9887-9897.
[11]	CAO J K, PANG J M, WENG X S, et al. Observation-centric SORT: rethinking sort for robust multi-object tracking[C]// 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2023: 9686-9696.
[12]	MAGLO A, ORCESI A, PHAM Q C. Efficient tracking of team sport players with few game-specific annotations[C]// 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. New York: IEEE Press, 2022: 3460-3470.
[13]	VATS K, WALTERS P, FANI M, et al. Player tracking and identification in ice hockey[J]. Expert Systems with Applications, 2023, 213: 119250. DOI URL
[14]	CUI Y T, JIANG C, WANG L M, et al. Mixformer: end-to-end tracking with iterative mixed attention[C]// 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2022: 13598-13608.
[15]	ZHANG Y F, SUN P Z, JIANG Y, et al. Bytetrack: multi-object tracking by associating every detection box[C]// European Conference on Computer Vision. Cham: Springer, 2022: 1-21.
[16]	HUANG H W, YANG C Y, SUN J C, et al. Iterative scale-up expansioniou and deep features association for multi-object tracking in sports[C]//2024 IEEE/CVF Winter Conference on Applications of Computer Vision Workshops. New York: IEEE Press, 2024: 163-172.
[17]	GRAN-HENRIKSEN M, LINDGAARD H A, KISS G, et al. Deep HM-SORT: enhancing multi-object tracking in sports with deep features, harmonic mean, and expansion IOU[EB/OL]. (2024-06-17) [2025-09-30]. https://arxiv.org/abs/2406.12081.
[18]	ZHOU K Y, YANG Y X, CAVALLARO A, et al. Omni-scale feature learning for person re-identification[C]// 2019 IEEE/ CVF International Conference on Computer Vision. New York: IEEE Press, 2019: 3701-3711.
[19]	WOJKE N, BEWLEY A, PAULUS D. Simple online and realtime tracking with a deep association metric[C]// 2017 IEEE International Conference on Image Processing. New York: IEEE Press, 2017: 3645-3649.
[20]	AHARON N, ORFAIG R, BOBROVSKY B Z. BoT-SORT: robust associations multi-pedestrian tracking[EB/OL]. (2022-07-07) [2025-09-28]. https://arxiv.org/abs/2206.14651.
[21]	ZHANG Y T, WANG S H, FAN Y X, et al. Translink: Transformer-based embedding for tracklets’ global link[C]// ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing. New York: IEEE Press, 2023: 1-5.
[22]	XIAO C C, CAO Q, LUO Z G, et al. Mambatrack: a simple baseline for multiple object tracking with state space model[C]// The 32nd ACM International Conference on Multimedia. New York: ACM, 2024: 4082-4091.
[23]	ESTER M, KRIEGEL H P, SANDER J, et al. A density-based algorithm for discovering clusters in large spatial databases with noise[C]// The 2nd International Conference on Knowledge Discovery and Data Mining. Washington DC: AAAI Press, 1996: 226-231.
[24]	ZHENG L, SHEN L Y, TIAN L, et al. Scalable person re-identification: a benchmark[C]// 2015 IEEE International Conference on Computer Vision. New York: IEEE Press, 2015: 1116-1124.
[25]	LUITEN J, OS̆EP A, DENDORFER P, et al. Hota: a higher order metric for evaluating multi-object tracking[J]. International Journal of Computer Vision, 2021, 129(2): 548-578. DOI PMID

运动类型	平均轨迹数	平均帧数	平均轨迹长度	平均轨迹密度
篮球	10	845.4	767.9	9.1
足球	22	673.9	422.1	12.8
排球	12	360.4	335.9	11.2

运动类型	平均轨迹数	平均帧数	平均轨迹长度	平均轨迹密度
篮球	10	845.4	767.9	9.1
足球	22	673.9	422.1	12.8
排球	12	360.4	335.9	11.2

方法	HOTA	AssA	MOTA	DetA
FairMOT	57.9	50.5	83.6	66.6
ByteTrack	71.5	60.7	94.6	84.3
BoT-SORT	77.0	63.4	95.6	93.5
OC-SORT	78.1	64.7	96.4	94.3
Deep-EIoU	85.4	73.6	96.5	99.2
DeepSORT	69.5	58.7	94.8	82.6
McByte	85.0	72.3	96.8	96.7
改进Deep-EIoU	89.1	79.4	96.7	99.1

方法	HOTA	AssA	MOTA	DetA
FairMOT	57.9	50.5	83.6	66.6
ByteTrack	71.5	60.7	94.6	84.3
BoT-SORT	77.0	63.4	95.6	93.5
OC-SORT	78.1	64.7	96.4	94.3
Deep-EIoU	85.4	73.6	96.5	99.2
DeepSORT	69.5	58.7	94.8	82.6
McByte	85.0	72.3	96.8	96.7
改进Deep-EIoU	89.1	79.4	96.7	99.1

方法	训练配置	HOTA/%	AssA/%	MOTA/%	DetA/%	IDF1/%	IDs
FairMOT	Train	49.3	34.7	86.4	70.2	53.5	9 928
CenterTrack	Train	62.7	48.0	90.8	82.1	60.0	10 481
ByteTrack	Train	62.8	51.2	94.1	77.1	69.8	3 267
ByteTrack	Train+Val	64.1	52.3	95.9	78.5	71.4	3 089
BoT-SORT	Train	68.7	55.9	94.5	84.4	70.0	6 729
OC-SORT	Train	71.9	59.8	94.5	86.4	72.2	3 093
OC-SORT	Train+Val	73.7	61.5	96.5	88.5	74.0	2 728
Deep-EIoU	Train	74.1	63.1	95.1	87.2	75.0	3 066
Deep-EIoU	Train+Val	77.2	67.7	96.3	88.2	76.8	2 659
QDTrack	Train	60.4	47.2	90.1	77.5	62.3	6 377
GTR	Train	54.5	45.9	67.9	64.8	55.8	9 567
TransTrack	Train	68.9	57.5	92.6	82.7	71.5	4 992
MixSort-Byte	Train+Val	65.7	54.8	96.2	78.8	74.1	2 472
MixSort-OC	Train+Val	74.1	62.0	96.5	88.5	74.4	2 781
MotionTrack	Train	74.0	61.7	96.6	88.8	74.0	—
DiffMOT	Train	72.1	60.5	94.5	86.0	72.8	—
DiffMOT	Train+Val	76.2	65.1	97.1	89.3	76.1	—
MambaTrack+	Train	71.3	58.6	94.9	86.7	71.1	—
MambaTrack	Train	72.6	60.3	95.3	87.6	72.8	—
ByteSSM	Train+Val	74.4	62.4	96.8	88.8	74.5	—
McByte	Train+Val	76.9	64.7	97.2	—	77.5	—
改进Deep-EIoU	Train	75.8	64.5	96.8	87.4	76.3	2 731
改进Deep-EIoU	Train+Val	78.1	68.4	97.1	88.6	80.1	2 538