1 College of Science, Shenyang Ligong University, Shenyang 110159, China
2 Liaoning Key Laboratory of Intelligent Optimization and Control for Ordnance Industry, Shenyang 110159, China
3 School of Automation and Electrical Engineering, Shenyang Ligong University, Shenyang 110159, China
4 National Key Laboratory of Electromagnetic Space Security, Tianjin 300308, China
Show less
文章历史+
Received
Published
Revised
2025-09-26
2026-10-18
2026-02-06
Online Date
2026-10-23
PDF (28728K)
摘要
针对红外场景下目标受遮挡导致跟踪身份频繁切换和轨迹中断问题,提出一种改进YOLOv12与StrongSORT的红外遮挡目标跟踪与计数算法AntiOcc-YOLOv12-StrongSORT。在检测阶段,基于YOLOv12框架引入排斥损失函数,通过增强红外遮挡目标的定位能力,提升相邻目标边界框的区分度,提高对遮挡目标的检测精度。在跟踪与计数阶段,建立遮挡感知机制,通过在StrongSORT框架中结合检测框的几何特征和目标的红外热辐射特征,判断目标遮挡状态,提高对遮挡目标的感知能力,缓解遮挡导致的目标身份丢失;设计抗遮挡动态模板更新策略,在目标外观模板更新过程中,依据目标遮挡状态自适应调整历史模板权重,抑制遮挡对目标外观模板的累积误差。基于跟踪结果引入虚拟区域计数策略,对红外遮挡目标进行精准计数。实验结果表明,与以YOLOv12为检测算法,分别融合Bytetrack、BoostSORT、DeepOCSORT、StrongSORT等跟踪器的多种融合跟踪算法相比,AntiOcc-YOLOv12-StrongSORT算法在多目标跟踪准确度(multiple object tracking accuracy,MOTA)、多目标跟踪精确度(multiple object tracking precision,MOTP)、ID调和均值(identification F1 score,IDF1)和身份转换数(identification switches,IDs)等跟踪指标均获得最优值,而且在不同复杂遮挡红外场景下均能够实现持续、稳定计数。
Abstract
Aiming at the problem of frequent identity switching and trajectory interruption caused by occluded targets in infrared scenarios, an improved YOLOv12 and StrongSORT-based algorithm named AntiOcc-YOLOv12-StrongSORT was proposed for infrared occluded object tracking and counting. During the detection stage, a repulsion loss function was integrated into YOLOv12 to enhance localization and distinguish adjacent bounding boxes, improving detection accuracy under occlusion. During the tracking and counting stage, an occlusion-aware mechanism was proposed within the StrongSORT framework. The occlusion state of targets was determined by combining the geometric features of detection boxes and the target’s infrared thermal radiation characteristics, thereby enhancing the perception of occluded targets and effectively mitigating identity loss caused by occlusion. Anti-occlusion dynamic template update strategy was proposed. During the update of the target appearance template, the weights of historical templates were adaptively adjusted based on the target’s occlusion state, significantly suppressing the cumulative error caused by occlusion in the target appearance template. A virtual region counting strategy was introduced based on tracking results to achieve accurate counting of infrared occluded targets. Experimental results indicate that compared to various tracking algorithms integrating YOLOv12 as the detection algorithm, such as Bytetrack, BoostSORT, DeepOCSORT, and StrongSORT, the AntiOcc-YOLOv12-StrongSORT algorithm achieves optimal values in multiple tracking metrics, including multiple object tracking accuracy (MOTA), multiple object tracking precision (MOTP), identification F1 score(IDF1), and identification switches(IDs). Furthermore, it demonstrates sustained and stable counting performance in complex occluded infrared scenarios.
LIU Fang, SUN Yi-qing, HU Yu-tong.
Algorithm for Tracking and Counting Infrared Occluded Targets Based on AntiOcc-YOLOv12-StrongSORT[J].
Science Technology and Engineering, 2026, 26(28): 12264-12276 DOI:10.12404/j.issn.1671-1815.2506916
近年来,深度学习方法凭借其强大的特征表征和学习能力,迅速成为目标跟踪领域的主导范式。根据算法网络结构不同,基于深度学习的目标跟踪方法主要分为联合检测的目标跟踪(joint detection and tracking,JDT)和基于检测的目标跟踪(tracking by detection,TBD)[5]。JDT目标跟踪方法通过单一网络同时输出目标检测结果和外观特征向量,实现检测与跟踪的并行处理,具有推理效率高的特点。但其性能受限于特征耦合,在区分外观高度相似的目标时易出现混淆,且对训练数据的规模和质量依赖较重。TBD目标跟踪方法则将跟踪任务解耦为独立的目标检测与目标跟踪两个阶段,通过先利用高性能检测器定位每帧中的目标,再基于运动模型和外观特征进行轨迹关联的方式对目标进行跟踪。该方法在复杂场景中具备更强的鲁棒性和更高的准确性,因此逐渐成为目标跟踪的主流算法框架。
由于TBD算法的性能与前端的检测结果和检测速度紧密关联,因此选择准确、快速的目标检测算法是保障目标跟踪稳定性、准确性和高效性的关键。依据算法处理的流程不同,基于深度学习的目标检测算法分为两阶段目标检测算法和单阶段目标检测算法[6],而单阶段目标检测算法凭借其高效性逐渐成为目标检测的主流方法。这类方法的核心是仅通过一个卷积神经网络(convolutional neural network,CNN)即可一体化完成目标类别判定与空间位置预测。其中,YOLO(you only look once)系列因其兼顾实时性与检测精度的特性,成为单阶段目标检测算法的核心代表。相关研究伴随着YOLO系列算法在持续优化检测精度与效率的过程中不断深入。同时,其算法在不同应用场景下的适用性进一步拓展,尤其在红外目标检测领域展现出广泛应用前景[7]。
基于YOLO系列的红外目标检测研究中,高永奇等[8]提出了一种基于改进YOLOv5的目标检测方法,通过引入InceptionNeXt作为主干网络,采用CARAFE(content-aware reassembly of features)算子替代特征融合层中上采样操作的策略,提高了水下废弃物红外目标检测的准确性和鲁棒性。许晓阳等[9]提出了一种改进的YOLOv7-tiny轻量级红外船舶目标检测算法,通过采用轻量级算法PP-LCNet,降低了网络参数量与计算量,验证了轻量级算法对红外目标的检测效果。Han等[10]基于YOLOv10框架提出一种多尺度特征融合改进的MFAF-YOLO,通过设计空间增强的高效多尺度注意力模块和跨尺度连接特征金字塔网络,提高特征的提取和表达能力,降低了在高速车辆超温区域检测中出现的漏检和误检数量。Ji等[11]提出了一种改进的YOLOv12检测框架,通过集成双向特征金字塔网络,进行多尺度特征融合,实现了输电线路的鲁棒缺陷检测,充分体现了YOLOv12在实际场景中的有效性与应用潜力。
在获得可靠的目标检测结果后,TBD框架的目标关联与轨迹维护是实现连续、稳定跟踪的关键。当前主流的跟踪算法主要包括以SORT[12] (simple online and realtime tracking)为代表的一系列多目标跟踪方法。其中,DeepSORT[13] (simple online realtime tracking with a deep association metric)、BotSORT[14](bottom-up tracking by sorting)、DeepOCSORT[15]和StrongSORT[16]等均为SORT演进方法,从关联度量、运动模型、外观特征融合等不同方面进行了改进,共同构成SORT系列的多目标跟踪体系。
YuanD, ZhangH, ShuX, et al. Thermal infrared target tracking: a comprehensive review[J]. IEEE Transactions on Instrumentation and Measurement, 2023, 73: 1-19.
[2]
MirzaeiB, Nezamabadi-PourH, RaoofA, et al. Small object detection and tracking: a comprehensive review[J]. Sensors, 2023, 23(15): 6887.
[3]
NosheenI, NaseerA, JalalA. 5thInternational Conference on Advancements in Computational Sciences ( ICACS).Efficient vehicle detection and tracking using blob detection and kernelized filter[C]// New York: IEEE, 2024: 1-8.
[4]
KumarA, VohraR, JainR, et al. Correlation filter based single object tracking: a review[J]. Information Fusion, 2024, 112: 102562.
LiYong, LinFang, ChenYuang, et al. A comprehensive review of deep learning-based visual multi-object tracking[J]. Science Technology and Engineering, 2025, 25(22): 9211-9223.
[7]
HuangJ, FuM, TangT, et al. Overview oftarget detection algorithm based on deep learning[C]// IEEE 2nd International Conference on Sensors, Electronics and Computer Engineering (ICSECE). New York: IEEE, 2024: 1174-1178.
[8]
WangL Y, BaiJ, LiW J, et al. Research progress of YOLO series target detection algorithms[J]. Journal of Computer Engineering & Applications, 2023, 59(14): 15-29.
XuXiaoyang, WeiWei, GaoChongyang. Infrared ship target detection algorithm based on improved YOLOv7-tiny[J]. Computer Engineering, 2026, 50(2): 209-220.
[13]
HanK, TangB, LiuD, et al. MFAF-YOLO: vehicle over-temperature region detection based on multi-scale feature attention fusion[J]. Measurement Science and Technology, 2025, 36(6): 065402.
[14]
JiY, MaT, ShenH, et al. Transmission line defect detection algorithm based on improved YOLOv12[J]. Electronics, 2025, 14(12): 2432.
[15]
BewleyA, GeZ, OttL, et al. Simple online and realtime tracking[C]// IEEE International Conference on Image Processing (ICIP). Phoenix, AZ: IEEE, 2016: 3464-3468.
[16]
WojkeN, BewleyA, PaulusD. Simple online and realtime tracking with a deep association metric[C]// IEEE International Conference on Image Processing (ICIP). Beijing: IEEE, 2017: 3645-3649.
MaggiolinoG, AhmadA, CaoJ, et al. DeepOCSORT: multi-pedestrian tracking by adaptive re-identification[C]// IEEE International Conference on Image Processing (ICIP). New York: IEEE, 2023: 3025-3029.
[19]
DuY, ZhaoZ, SongY, et al. StrongSORT: Make DeepSORT great again[J]. IEEE Transactions on Multimedia, 2023, 25: 8725-8737.
LiuZhixi. Research on improved YOLOv5 and DeepSORT infrared multi-target detection and tracking methods[D]. Taiyuan: North University of China, 2024.
[22]
ChenY H. Strong Baseline: Multi-UAVtracking via YOLOv12 with BoT-SORT-ReID[C]// Proceedings of the Computer Vision and Pattern Recognition Conference. Nashville: IEEE, 2025: 6573-6582.
[23]
ZhaC, LuoS, XuX. Infrared multi-target detection and tracking in dense urban traffic scenes[J]. IET Image Processing, 2024, 18(6): 1613-1628.
[24]
AlbouchiA, MessaoudS, BouaafiaS, et al. Implementation of an improved multi-object detection, tracking, and counting for autonomous driving[J]. Multimedia Tools and Applications, 2024, 83(18): 53467-53495.
ShenYang, WangChongyu, ZhaoJiayi, et al. Fish fry counting method based on improved YOLOv8 and multi-target tracking[J]. Transactions of the Chinese Society of Agricultural Engineering, 2024, 40(16): 163-170.
AnZhiqiang, LiZhijun, LiuShuo, et al. Multi-category sea treasure counting method integrating YOLOv7 and BYTE multi-target tracking[J]. Transactions of the Chinese Society of Agricultural Engineering, 2023, 39(9): 183-189.
LüJia, ZhangCuiping. Real-time detection and counting of double-associated bagged grape clusters based on EMO-YOLOv5s[J]. Transactions of the Chinese Society of Agricultural Engineering, 2025, 41(12): 161-171.
[31]
WangX, XiaoT, JiangY, et al. Repulsion loss: detecting pedestrians in a crowd[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. New York: IEEE, 2018: 7774-7783.
[32]
GilroyS, GlavinM, JonesE, et al. An objective method for pedestrian occlusion level classification[J]. Pattern Recognition Letters, 2022, 164: 96-103.