基于动态特征的海上弹着点水柱检测跟踪算法

doi:10.12382/bgxb.2024.0918

摘要/Abstract

摘要：

利用可见光图像对海上弹着点水柱进行有效检测与跟踪,是实现海上自动检靶的核心。由于相机的运动、焦距的调节以及水柱的变化,现有的检测与跟踪算法依然存在较高的误报率和身份切换次数(Identity Switch Times,IDs)。为解决上述问题,提出一种基于动态特征的海上弹着点水柱检测跟踪算法。利用YOLOv8目标检测器对静态水柱进行检测,通过在浅层特征图上增加小目标检测头,增强模型对小水柱的检测能力;利用改进的ByteTrack跟踪器对水柱进行跟踪,将相机运动和卡尔曼滤波相结合,补偿由相机运动引起的跟踪偏移;结合水柱形成阶段的时空特征,采用支持向量机进行综合决策,实现对水柱的判断。实验结果表明,与传统的检测跟踪算法相比,新算法在多目标跟踪准确度、识别平均数比率和多目标跟踪精确度这3个关键性能指标上分别提升了7.8%、5.1%和0.9%;误报数减少了112次,IDs数和误检数均降至0,表明新算法不仅能够精确地检测和跟踪水柱,还能够有效地排除其他干扰因素,在整体性能上实现显著的增强。

关键词: 动态特征, 运动估计, 目标检测, 目标跟踪, ByteTrack

Abstract:

The effective detection and tracking of water columns at marine impact points using visible light images is key to automatically check a target at sea. The existing detection and tracking algorithms still have a high false alarm rate and identity switch times (IDs) due to the movement of camera, the adjustment of focal length, and the changes of water columns. To solve the above problems, this paper proposes a detection and tracking algorithm based on dynamic features for water columns at marine impact points. The YOLOv8 target detector is used to detect the static water columns, and a small target detection head is added to the shallow feature map to enhance the model’s ability to detect small water columns. An improved ByteTrack tracker is used to track the water columns, and the tracking offsets caused by camera movement is compensated by combining camera movement and Kalman filtering. And then, a support vector machine is used for comprehensive decision-making to judge the water columns according to the spatiotemporal features of the water columns formation stage. Compared with traditional detection and tracking algorithms, the proposed algorithm is used to improve the three key performance indicators of multiple object tracking accuracy (MOTA), identification F1 (IDF1), and multiple object tracking precision (MOTP) by 7.8%, 5.1%, and 0.9%, respectively, the number of false positives (FP) is reduced by 112 times, and the numbers of IDs and false detections are both reduced to zero. Experimental results show that the proposed algorithm can not only accurately detect and track the water columns but also effectively exclude other interfering factors, thus achieving a significant enhancement in overall performance.

Key words: dynamic feature, motion estimation, target detection, target tracking, ByteTrack

中图分类号:

TJ301

桂凡, 石章松, 孙世岩, 应文健, 胡卫强, 徐慧慧, 吴中红, 胡清平, 张俊. 基于动态特征的海上弹着点水柱检测跟踪算法[J]. 兵工学报, 2025, 46(5): 240918-.

GUI Fan, SHI Zhangsong, SUN Shiyan, YING Wenjian, HU Weiqiang, XU Huihui, WU Zhonghong, HU Qingping, ZHANG Jun. Detection and Tracking of Water Columns at Marine Impact Points Based on Dynamic Features[J]. Acta Armamentarii, 2025, 46(5): 240918-.

图/表 13

图1 基于动态特征的水柱检测跟踪算法框架

Fig.1 Detection and tracking algorithm framework for water columns based on dynamic features

图2 YOLOv8-small检测模型结构图

Fig.2 YOLOv8-small detection model structure diagram

图3 由CME进行水柱检测框映射

Fig.3 Water column detection box mapping based on CME

图4 目标面积随时间变化曲线

Fig.4 Changing curve of target area over time

图5 目标平均增长率随序列数的变化

Fig.5 Variation of target average growth rate with sequence number

表1 测试集

Table 1 Test set

视频序号	视角	分辨率	帧率	时长	水柱个数
001	水面	640×352	21	3min18s	21
002	空中	1920×1080	30	2min25s	10
003	空中	1920×1080	50	2min59s	10

表2 模型算法性能

Table 2 Performance comparison of the algorithm

算法	P/%	R/%	mAP50/%	mAP50-95/%
YOLOv8	0.939	0.887	0.948	0.799
YOLOv8-samll	0.953	0.915	0.973	0.817

表3 改进前后检测结果效果对比

Table 3 Comparison of detected results before and after improvement

表4 基于静态特征的水柱检测跟踪结果

Table 4 Detection and tracking results of water columns based on static features

视频序号与均值	IDF1/%	GT	MT	FP	FN	IDs	MOTA/%	MOTP/%	检靶数	误检数	漏检数
001	98.4	21	21	5	7	0	96.8	19.7	21	5	0
002	90.2	10	10	75	1	5	88.9	7.0	10	7	0
003	91.4	10	10	56	4	0	81.3	7.5	10	11	0
均值	93.3						89	11.4

表5 基于动态特征的水柱检测跟踪结果

Table 5 Detection and tracking results of water columns based on dynamic features

视频序号及均值	IDF1/%	GT	MT	FP	FN	IDs	MOTA/%	MOTP/%	检靶数	误检数	漏检数
001	98.8	21	21	2	7	0	97.6	19.7	21	0	0
002	98.3	10	10	12	13	0	96.6	9.9	10	0	0
003	98.1	10	10	8	4	0	96.3	7.5	10	0	0
均值	98.4						96.8	12.3

表6 测试视频001检测跟踪效果对比

Table 6 Comparison of tracking effects for test video 001

表7 测试视频002检测跟踪效果对比

Table 7 Comparison of tracking effects for test video 002

表8 测试视频003检测跟踪效果对比

Table 8 Comparison of tracking effects for test video 003

参考文献 28

[1]	王永生, 姬嗣愚, 杜彬彬. 基于改进Faster R-CNN的海上弹着点水柱目标检测算法[J]. 兵器装备工程学报, 2022, 43(6):182-189.
	WANG Y S, JI S Y, DU B B. Marine impact water column signal detection algorithm based on improved Faster R-CNN[J]. Journal of Ordnance Equipment Engineering, 2022, 43(6): 182-189. (in Chinese)
[2]	张坤, 罗亚松, 刘忠. 基于YOLOv4的海上目标识别技术研究[J]. 兵器装备工程学报, 2022, 43(4): 211-217.
	ZHANG K, LUO Y S, LIU Z. Research on marine target recognition technology based on YOLOv4[J]. Journal of Ordnance Equipment Engineering, 2022, 43(4):211-217. (in Chinese)
[3]	姬嗣愚, 王永生. 基于改进YOLOv5的海上弹着点水柱信号检测算法[J]. 电光与控制, 2022, 29(8):50-56.
	JI S Y, WANG Y S. An improved YOLOv5 based algorithm for water column signal detection at marine impact point[J]. Electronics Optics & Control, 2022, 29(8):50-56. (in Chinese)
[4]	杜立召, 徐岩, 张为. 一种双网融合的分阶段烟雾检测算法[J]. 西安电子科技大学学报(自然科学版), 2020, 47(4):141-148.
	DU L Z, XU Y, ZHANG W. Phased smoke detection algorithm using dual network fusion[J]. Journal of Xidian University, 2020, 47(4):141-148. (in Chinese)
[5]	郑怀兵, 翟济云. 基于视频分析的森林火灾烟雾检测方法[J]. 南京理工大学学报(自然科学版), 2015, 39(6): 686-691,710.
	ZHENG H B, ZHAI J Y. Forest fire smoke detection based on video analysis[J]. Journal of Nanjing University of Science and Technology, 2015, 39(6):686-691,710. (in Chinese)
[6]	赵敏, 张为, 王鑫, 等. 时空背景模型下结合多种纹理特征的烟雾检测[J]. 西安交通大学学报, 2018, 52(8): 67-73.
	ZHAO M, ZHANG W, WANG X, et al. Smoke detection algorithm multi-texture feature exploration under a spatio-temporal background model[J]. Journal of Xi’an Jiaotong University, 2018, 52(8):67-73. (in Chinese)
[7]	李笋, 石永生, 汪渤, 等. 基于颜色增强变换和MSER检测的烟雾检测算法[J]. 北京理工大学学报, 2016, 36(10):1072-1078.
	LI S, SHI Y S, WANG B, et al. Video smoke detection based on color transformation and MSER[J]. Transactions of Beijing Institute of Technology, 2016, 36(10):1072-1078. (in Chinese)
[8]	刘通, 程江华, 华宏虎, 等. 结合Y^dU^aV^a颜色模型和改进MobileNetV3的视频烟雾检测方法[J]. 国防科技大学学报, 2021, 43(5):80-85.
	LIU T, CHENG J H, HUA H H, et al. Video smoke detection method combined Y^dU^aV^a color model and improved MobileNetV3[J]. Journal of National University of Defense Technology, 2021, 43(5):80-85. (in Chinese)
[9]	刘志赢, 谢春思, 李进军, 等. 基于改进Deeplabv3+的烟雾区域分割识别算法[J]. 系统工程与电子技术, 2021, 43(2):328-335. doi: 10.12305/j.issn.1001-506X.2021.02.06
	LIU Z Y, XIE C S, LI J J, et al. Smoke region segmentation recognition algorithm based on improved Deeplabv3+[J]. Systems Engineering and Electronics, 2021, 43(2):328-335. (in Chinese) doi: 10.12305/j.issn.1001-506X.2021.02.06
[10]	王浩远, 梁煜, 张为. 融合多分辨率表征的实时烟雾分割算法[J]. 浙江大学学报(工学版), 2021, 55(12):2334-2341.
	WANG H Y, LIANG Y, ZHANG W. Real-time smoke segmentation algorithm fused with multi-resolution repres-entation[J]. Journal of Zhejiang University (Engineering Science), 2021, 55(12):2334-2341. (in Chinese)
[11]	HU Y W, ZHAN J L, ZHOU G X, et al. Fast forest fire smoke detection using MVMNet[J]. Knowledge-Based Systems, 2022, 241:108219.
[12]	高武奇, 杨婷, 李亮亮. 基于多特征交叉融合及跨层级联的航拍目标检测算法[J]. 西北工业大学学报, 2023, 41(6):1179-1189.
	GAO W Q, YANG T, LI L L. Aerial military target detection algorithm based on multi-feature cross fusion and cross-layer concatenation[J]. Journal of Northwestern Polytechnical University, 2023, 41(6):1179-1189. (in Chinese)
[13]	马庆禄, 鲁佳萍, 唐小垚, 等. 改进YOLOv5s的公路隧道烟火检测方法[J]. 浙江大学学报(工学版), 2023, 57(4): 784-794,813.
	MA Q L, LU J P, TANG X Y, et al. Improved YOLOv5s flame and smoke detection method in road tunnels[J]. Journal of Zhejiang University(Engineering Science), 2023, 57(4):784-794,813. (in Chinese)
[14]	王殿伟, 赵文博, 房杰, 等. 多类场景下无人机航拍视频烟雾检测算法[J]. 哈尔滨工业大学学报, 2023, 55(10): 122-129.
	WANG D W, ZHAO W B, FANG J, et al. Smoke detection algorithm for UAV aerial video in multiple scenarios[J]. Journal of Harbin Institute of Technology, 2023, 55(10): 122-129. (in Chinese)
[15]	XU Z, DAI T Q. Efficient fire-smoke detection and evacuation simulation from buildings based on YoLoX-Swin[J]. Journal of Southeast University, 2023, 39(4): 372-383.
[16]	QIAO Y M, JIANG W Y, WANG F, et al. FireFormer: an efficient transformer to identify forest fire from surveillance cameras[J]. International Journal of Wildland Fire, 2023, 32(9): 1364-1380.
[17]	CHEN Y S, LI J H, SUN K, et al. A lightweight early forest fire and smoke detection method[J]. The Journal of Supercomputing, 2024, 80(7): 9870-9893.
[18]	YIN D X, CHENG P L, HUANG Y. YOLO-EPF: multi-scale smoke detection with enhanced pool former and multiple receptive fields[J]. Digital Signal Processing, 2024, 149: 104511.
[19]	陈俊周, 汪子杰, 陈洪瀚, 等. 基于级联卷积神经网络的视频动态烟雾检测[J]. 电子科技大学学报, 2016, 45(6):992-996.
	CHEN J Z, WANG Z J, CHEN H H, et al. Dynamic smoke detection using cascaded convolutional neural network for surveillance videos[J]. Journal of University of Electronic Science and Technology of China, 2016, 45(6): 992-996. (in Chinese)
[20]	SIMONYAN K, ZISSERMAN A. Two-stream convolutio-nal networks for action recognition in videos[J]. Advances in Neural Information Processing Systems, 2014, 27: 568-576.
[21]	LI C Y, LI L L, JIANG H L, et al. YOLOv6: a single-stage object detection framework for industrial applications: arXiv:2209.02976[R]. Ithaca, NY, US: Cornell University, 2022: 2209.02976.
[22]	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:arXiv:2207. 02696[R]. Ithaca,NY, US: Cornell University,2022:2207.02696.
[23]	WANG C Y, YEH I H, LIAO H Y M. YOLOv9: learning what you want to learn using programmable gradient Information: arXiv: 2402.13616[R]. Ithaca, NY. US: Cornell University, 2024: 2402.13616.
[24]	WANG A, CHEN H, LIU L H, et al. YOLOv10:real-time end-to-end object detection: arXiv: 2405.14458[4]. Ithaca, NY. US: Cornell University, 2024: 2405.14458.
[25]	ZHANG Y F, SUN P Z, JIANG Y, et al. ByteTrack: multi-object tracking by associating every detection box: arXiv:2110.06864[R]. Ithaca, NY. US: Cornell University, 2021: 2110.06864.
[26]	郭爱强, 李天鹏, 朱曦, 等. 基于卷积神经网络和时空特征的烟幕视频检测和参数提取[J]. 兵工学报, 2024, 45(8): 2478-2486. doi: 10.12382/bgxb.2023.0595
	GUO A Q, LI T P, ZHU X, et al. Smoke screen video detection and parameter extraction based on convolutional neural network and spatio-temporal features[J]. Acta Armamentarii, 2024, 45(8):2478-2486. (in Chinese) doi: 10.12382/bgxb.2023.0595
[27]	CALDERARA S, PICCININI P, CUCCHIARA R. Vision based smoke detection system using image energy and Color Information[J]. Machine Vision and Applications, 2011, 22(4): 705-719.
[28]	DENDORFER P, REZATOFIGHI H, MILAN A, et al. MOT20: a benchmark for multi object tracking in crowded scenes: arXiv:2003.09003[R]. Ithaca, NY. US: Cornell University, 2020: 2003.09003.