面向智能船艇航行场景的小目标高精度跟踪器

邵泽远; 尹勇; 吕红光; 景乾峰; 王海超

doi:10.3963/j.jssn.1674-4861.2025.04.008

面向智能船艇航行场景的小目标高精度跟踪器

doi: 10.3963/j.jssn.1674-4861.2025.04.008

1.
大连海事大学航海动态仿真和控制交通部重点实验室辽宁大连 116026
2.
广州航海学院航运学院广州 510725

基金项目:

国家重点研发计划项目 2022YFB4300803

国家重点研发计划项目 2022YFB4301402

国家自然科学基金项目 52071049

辽宁省自然科学基金博士科研启动基金计划项目 2024-BS-013

大连市揭榜挂帅技术攻关项目 2024JB11PT007

详细信息

作者简介:
邵泽远（1996—），博士研究生. 研究方向：智能船导航系统、海事视觉信息感知. E-mail：szy@dlmu.edu.cn

通讯作者:
吕红光（1981—），博士，教授. 研究方向：海上智能交通系统、船舶自动避碰及路径规划等. E-mail：lhg@dlmu.edu.cn

中图分类号: U675.73
计量
- 文章访问数: 2
- HTML全文浏览量: 2
- PDF下载量: 0
- 被引次数: 0
出版历程
- 收稿日期: 2025-03-07

A High-precision Tracker for Small Objects in Intelligent Vessel Navigation Scenarios

1.
Key Laboratory of Marine Simulation & Control for Ministry of Transportation Dalian Maritime University, Dalian 116026, Liaoning, China
2.
School of Shipping and Maritime Studies Guangzhou Maritime University, Guangzhou 510725, China

摘要

摘要: 海上小目标跟踪在智能船艇环境感知中至关重要。然而，受小目标特征信息少和船载相机抖动的影响，现有方法在跟踪精度与稳定性方面仍存在不足。为提升海上小目标的跟踪性能，研究了1种新颖的海上小目标跟踪器（SeaMicroTracker）。该跟踪器将基于深度学习的目标检测技术与优化的目标关联算法相结合，从而实现鲁棒的海上小目标跟踪。具体来说，目标检测部分采用增强型卷积神经网络（eYOLOv5）模型，以精确获取海上小目标的位置信息。目标关联部分中，设计了以观测为中心的卡尔曼滤波方法（observation-centric Kalman filter，OKF），以提高船载相机运动条件下状态估计的可靠性；提出了1种基于曼哈顿距离的交并比（Manhattan distance intersection over union，MDIoU）度量方式，以提升海上动态环境下的关联精度；并设计了1种基于逐级细化的级联匹配（progressive refinement cascade matching，PRCM）策略，以增强跟踪器对海上复杂环境变化、目标遮挡干扰的适应能力，进一步提升了海上目标关联能力。在杰瑞海事跟踪数据集上的实验结果表明，SeaMicroTracker在跟踪准确度和稳定性方面具有显著优势，多目标跟踪准确度（multiple object tracking accuracy，MOTA）和ID匹配得分（identification F1，IDF1）分别为80.6和64.0。与基线方法ByteTrack相比，所提方法的MOTA和IDF1分数分别提升了27.9%和34.7%，有效减少了ID切换次数，且平均跟踪速率达30.1FPS，满足工程应用需求。
- 交通安全 /
- 智能船艇 /
- 海上小目标跟踪 /
- 深度学习 /
- 卷积神经网络 /
- 卡尔曼滤波
Abstract: Tracking small sea-surface objects is crucial for environmental perception in intelligent vessels. However, due to the limited feature information of small objects and the motion-induced instability of shipborne cameras, existing methods still face challenges in tracking accuracy and stability. To address these issues and enhance the performance of small sea-surface objects tracking, a novel tracker named SeaMicroTracker is developed. The tracker integrates deep learning-based object detection with an optimized object association algorithm to enhance robustness. Specifically, for object detection, an enhanced Convolutional Neural Network (eYOLOv5) model is employed to accurately extract the positional information of small sea-surface objects. For object association, an observation-centric Kalman filter (OKF) is designed to enhance state estimation reliability under shipborne camera motion. Additionally, a Manhattan distance-based intersection over union (MDIoU) metric is proposed to improve association accuracy in dynamic maritime environments. Furthermore, a progressive refinement cascade matching (PRCM) strategy is developed to improve the tracker's adaptability to complex maritime conditions and target occlusions, thereby further enhancing target association performance. Experimental results on the Jari maritime tracking dataset show that SeaMicroTracker achieves a multiple object tracking accuracy (MOTA) of 80.6 and an identification F1 score (IDF1) of 64.0, demonstrating significant advantages in tracking accuracy and stability. Compared with the baseline method ByteTrack, the proposed tracker improves MOTA and IDF1 by 27.9% and 34.7%, respectively, while effectively reducing ID-switching events. Moreover, the tracker achieves an average tracking speed of 30.1 FPS, satisfying real-time requirements for engineering applications.
- traffic safety /
- intelligent vessels /
- small sea-surface object tracking /
- deep learning /
- convolutional neural network /
- Kalman filter

HTML全文

图 1 海上小目标跟踪器SeaMicroTracker的整体流程图

Figure 1. Workflow of the small sea-surface object tracker SeaMicroTracker

下载: 全尺寸图片幻灯片

图 2 eYOLOv5模型的网络架构

Figure 2. Network architecture of the eYOLOv5 model

下载: 全尺寸图片幻灯片

图 3 特征图可视化结果

Figure 3. Feature map visualization results

下载: 全尺寸图片幻灯片

图 4 数据集实例

Figure 4. Examples from the dataset

下载: 全尺寸图片幻灯片

图 5 标签中心点及大小分布

Figure 5. Distribution of label center points and sizes

下载: 全尺寸图片幻灯片

图 6 不同方法自由组合对跟踪准确度指标的影响

Figure 6. Impact of different method combinations on tracking accuracy metrics

下载: 全尺寸图片幻灯片

图 7 不同模块自由组合对检测准确度指标的影响

Figure 7. Impact of different module combinations on detection accuracy metrics

下载: 全尺寸图片幻灯片

图 8 关联算法结合各检测模型的跟踪性能可视化

Figure 8. Visualization of tracking performance for detection models combined with association algorithms

下载: 全尺寸图片幻灯片

图 9 各跟踪器在MOTA、IDF1、FPS上的性能分析

Figure 9. Trackers'Performance Analysis on MOTA, IDF1, and FPS

下载: 全尺寸图片幻灯片

图 10 本文跟踪器与基线方法在测试集上的小目标跟踪结果比较

Figure 10. Comparison of small object tracking results between the proposed tracker and baseline on the test set

下载: 全尺寸图片幻灯片

图 11 所提跟踪器在不同实海域视频片段中的跟踪效果

Figure 11. Tracking performance of the proposed tracker on different real-sea video clips

下载: 全尺寸图片幻灯片

图 12 SeaMicroTracker在渤海轮渡场景下的泛化性能验证结果

Figure 12. Generalization performance verification results of SeaMicroTracker in the Bohai ferry scene

下载: 全尺寸图片幻灯片

表 1 训练关键参数设置

Table 1. Key parameter settings for training

参数名称	设置
学习率(Learning Rate)	0.01
图像大小(Image Size)	640×640
优化器(Optimizer)	Adam
动量值(Momentum Value)	0.937
权重衰减(Weight Decay)	5×10^-4
批量大小(Batch Size)	24
训练时长(Training Epochs)	200

下载: 导出CSV

表 2 关联部分不同组件消融实验结果

Table 2. Ablation study results of different components in the association module

实验序列	BYTE	OKF	MDIoU	PRCM	MOTA ↑	IDF1 ↑
1	√	×	×	×	63.0	47.5
2	√	√	×	×	75.2	58.6
3	√	×	√	×	67.0	52.1
4	√	×	×	√	64.2	48.4
5	√	√	√	×	79.5	63.3
6	√	√	×	√	76.2	59.3
7	√	×	√	√	67.9	52.6
8	√	√	√	√	80.6	64.0

下载: 导出CSV

表 3 检测部分不同组件消融实验结果

Table 3. Ablation study results of different components in the detection module

实验序列	YOLOv5	PAN*	DCNv2	SCAM++	mAP_0.5:0.95↑	mAP_0.5↑
1	√	×	×	×	0.467	0.839
2	√	√	×	×	0.487	0.875
3	√	×	√	×	0.468	0.845
4	√	×	×	√	0.474	0.848
5	√	√	√	×	0.489	0.882
6	√	√	×	√	0.488	0.880
7	√	×	√	√	0.476	0.849
8	√	√	√	√	0.496	0.884

下载: 导出CSV

表 4 所提关联算法结合不同检测模型的性能比较

Table 4. Performance comparison of the proposed association algorithm with different detection models

检测模型	关联算法	MOTA ↑	IDF1 ↑
YOLOv5^[17]	√	78.3	60.3
YOLOv5^[17]	×	60.2	44.5
YOLOX^[18]	√	78.9	61.9
YOLOX^[18]	×	60.8	45.7
YOLOv7^[19]	√	79.4	62.2
YOLOv7^[19]	×	61.5	45.8
YOLOv8^[20]	√	79.7	62.3
YOLOv8^[20]	×	61.7	46.1
eYOLOv5^[25]	√	80.6	64.0
eYOLOv5^[25]	×	63.0	47.5

下载: 导出CSV

表 5 不同跟踪器性能比较

Table 5. Comparison of performance among different trackers

跟踪器	MOTA ↑	IDF1 ↑	FPS ↑
SORT^[21]	39.1	24.5	35.4
DeepSORT^[22]	27.9	26.8	15.8
ByteTrack^[23]	63.8	49.0	30.6
BoT-SORT^[24]	75.7	62.5	11.8
DeepMOT^[30]	38.2	37.9	16.8
UAVMOT^[31]	42.3	35.2	19.2
C-BIoU^[27]	68.8	46.7	26.5
SeaMicroTracker	80.6	64.0	30.1

下载: 导出CSV

参考文献(31)

[1]	娄月新, 金强, 魏明泽, 等. 海上自主水面船舶自主运行模式及未来发展分析[J]. 船海工程, 2025, 54(1): 71-75, 80. LOU Y X, JIN Q, WEI M Z, et al. Analysis of the autonomous operation mode and future development for maritime autonomous surface ship[J]. Ship & Ocean Engineering, 2025, 54(1): 71-75, 80. (in Chinese)
[2]	胡一鹏, 闫昭琨, 刘佳仑, 等. 智能船艇虚实融合测试验证技术现状与展望[J]. 船舶工程, 2022, 44(4): 4-13. HU Y P, YAN Z K, LIU J L, et al. Current status and prospects of virtual-real fusion test and verification technology for intelligent vessels[J]. Ship Engineering, 2022, 44(4): 4-13. (in Chinese)
[3]	THOMBRE S, ZHAO Z, RAMM-SCHMIDT H, et al. Sensors and AI techniques for situational awareness in autonomous ships: a review[J]. IEEE Transactions on Intelligent Transportation Systems, 2022, 23(1): 64-83. doi: 10.1109/TITS.2020.3023957
[4]	LYU H, SHAO Z, CHENG T, et al. Sea-surface object detection based on electro-optical sensors: a review[J]. IEEE Intelligent Transportation Systems Magazine, 2023, 15(2): 190-216. doi: 10.1109/MITS.2022.3198334
[5]	CHEN X, LING J, WANG S, et al. Ship detection from coastal surveillance videos via an ensemble Canny-Gaussian-morphology framework[J]. The Journal of Navigation, 2021, 74 (6): 1252-1266. doi: 10.1017/S0373463321000540
[6]	CHEN X, XU X, YANG Y, et al. Visual ship tracking via a hybrid kernelized correlation filter and anomaly cleansing framework[J]. Applied Ocean Research, 2021, 106: 102455. doi: 10.1016/j.apor.2020.102455
[7]	CHEN Z, CHEN D, ZHANG Y, et al. Deep learning for autonomous ship-oriented small ship detection[J]. Safety Science, 2020, 130: 104812. doi: 10.1016/j.ssci.2020.104812
[8]	LIU R W, YUAN W, CHEN X, et al. An enhanced CNN-enabled learning method for promoting ship detection in maritime surveillance system[J]. Ocean Engineering, 2021, 235: 109435. doi: 10.1016/j.oceaneng.2021.109435
[9]	SHAO Z, LYU H, YIN Y, et al. Multi-scale object detection model for autonomous ship navigation in maritime environment[J]. Journal of Marine Science and Engineering, 2022, 10 (11): 1783. doi: 10.3390/jmse10111783
[10]	王宁, 吴伟, 王元元, 等. 多特征融合的无人艇视觉目标长时相关鲁棒跟踪[J]. 中国舰船研究, 2024, 19(1): 62-74. WANG N, WU W, WANG Y Y, et al. Long-term correlation robust tracking of visual targets for unmanned surface vehicles using multi-feature fusion[J]. Chinese Journal of Ship Research, 2024, 19(1): 62-74. (in Chinese)
[11]	陈惠红, 胡耀民, 刘世明. 交互多模粒子滤波多特征自适应融合的船舶视觉跟踪[J]. 舰船科学技术, 2018, 40(4): 133-135. CHEN H H, HU Y M, LIU S M. Interactive multi-mode particle filter multi feature adaptive fusion for ship vision tracking[J]. Ship Science and Technology, 2018, 40(4): 133-135. (in Chinese)
[12]	BLOISI D, IOCCHI L, FIORINI M, et al. Automatic maritime surveillance with visual target detection[C]. The International Defense and Homeland Security Simulation Workshop, Rome, Italy: University of Genoa, 2011.
[13]	CHEN X, WANG S, SHI C, et al. Robust ship tracking via multi-view learning and sparse representation[J]. The Journal of Navigation, 2019, 72(1): 176-192. doi: 10.1017/S0373463318000504
[14]	TIAN W, LAUER M, CHEN L. Online multi-object tracking using joint domain information in traffic scenarios[J]. IEEE Transactions on Intelligent Transportation Systems, 2019, 21 (1): 374-384.
[15]	郭晓晗, 彭理群, 马定辉. 基于车联网BSM数据与路侧视频融合的港口集装箱卡车碰撞危险辨识方法[J]. 交通信息与安全, 2023, 41(1): 1-12. doi: 10.3963/j.jssn.1674-4861.2023.01.001 GUO X H, PENG L Q, MA D H. A method of identifying collision risk of container trucks in port terminal areas under an integrated connected vehicle BSM and roadside video surveillance data[J]. Journal of Transport Information and Safety, 2023, 41(1): 1-12. (in Chinese) doi: 10.3963/j.jssn.1674-4861.2023.01.001
[16]	HASSAN S, MUJTABA G, RAJPUT A, et al. Multi-object tracking: a systematic literature review[J]. Multimedia Tools and Applications, 2024, 83(14): 43439-43492.
[17]	马浩为, 张笛, 李玉立, 等. 基于改进YOLOv5的雾霾环境下船舶红外图像检测算法[J]. 交通信息与安全, 2023, 41 (1): 95-104. doi: 10.3963/j.jssn.1674-4861.2023.01.010 MA H W, ZHANG D, LI Y L, et al. A ship detection algorithm for infrared images under hazy environment based on an improved YOLOv5 algorithm[J]. Journal of Transport Information and Safety, 2023, 41(1): 95-104. (in Chinese) doi: 10.3963/j.jssn.1674-4861.2023.01.010
[18]	张立国, 赵嘉士, 金梅, 等. 改进YOLOX在近岸船舶检测中的应用[J]. 计量学报, 2024, 45(1): 30-37. ZHANG L G, ZHAO J S, JIN M, et al. Application of improved YOLOX in inshore ship inspection[J]. Acta Metrologica Sinica, 2024, 45(1): 30-37. (in Chinese)
[19]	WANG C Y, BOCHKOVSKIY A, LIAO H Y M. YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors[C]. The IEEE Conference on Computer Vision and Pattern Recognition, Vancouver, Canada: IEEE, 2023.
[20]	周欣, 徐培哲, 李堃, 等. 基于损失函数与注意力机制改进的YOLOv8火焰目标检测算法优化研究[J]. 船海工程, 2025, 54(2): 19-25. ZHOU X, XU P Z, LI K, et al. Optimization of YOLOv8 flame target detection algorithm based on improved loss function and attention mechanism[J]. Ship & Ocean Engineering, 2025, 54(2): 19-25. (in Chinese)
[21]	BEWLEY A, GE Z, OTT L, et al. Simple online and realtime tracking[C]. The IEEE International Conference on Image Processing, Phoenix, USA: IEEE, 2016.
[22]	WOJKE N, BEWLEY A, PAULUS D. Simple online and realtime tracking with a deep association metric[C]. The IEEE International Conference on Image Processing, Beijing, China: IEEE, 2017.
[23]	ZHANG Y, SUN P, JIANG Y, et al. Bytetrack: multi-object tracking by associating every detection box[C]. 17^th European Conference on Computer Vision, Tel Aviv, Israel: Springer Nature Switzerland, 2022.
[24]	AHARON N, ORFAIG R, BOBROVSKY B Z. BoT-SORT: robust associations multi-pedestrian tracking[EB/OL]. (2022-07-07)[2025-03-01]. https://arxiv.org/abs/2206.14651.
[25]	SHAO Z, YIN Y, LYU H, et al. An efficient model for small object detection in the maritime environment[J]. Applied Ocean Research, 2024, 152: 104194. doi: 10.1016/j.apor.2024.104194
[26]	ZHENG Z, WANG P, REN D, et al. Enhancing geometric factors in model learning and inference for object detection and instance segmentation[J]. IEEE Transactions on Cybernetics, 2021, 52(8): 8574-8586.
[27]	YANG F, ODASHIMA S, MASUI S, et al. Hard to track objects with irregular motions and similar appearances? make it easier by buffering the matching space[C]. The IEEE Winter Conference on Applications of Computer Vision, Waikoloa, USA: IEEE, 2023.
[28]	Jiangsu Automation Research Institute. Jari maritime tracking dataset. [DB/OL]. (2022-09-30)[2025-03-01]. https://github.com/hjq3659/Jari-Maritime-Tracking-Dataset.git.
[29]	LIN T Y, MAIRE M, BELONGIE S, et al. Microsoft COCO: common objects in context[C]. 13^th European Conference on Computer Vision, Zurich, Switzerland: Springer Nature Switzerland, 2014.
[30]	XU Y, OSEP A, BAN Y, et al. How to train your deep multi-object tracker[C]. The IEEE Conference on Computer Vision and Pattern Recognition, Seattle, USA: IEEE, 2020.
[31]	LIU S, LI X, LU H, et al. Multi-object tracking meets moving UAV[C]. The IEEE Conference on Computer Vision and Pattern Recognition, New Orleans, USA: IEEE, 2022.