水下弱目标跟踪的深度学习方法研究

doi:10.12382/bgxb.2022.0842

摘要/Abstract

摘要：

针对水下弱目标跟踪受干扰和噪声影响容易出现量测丢失或偏差,导致传统Kalman滤波方法跟踪误差显著增加甚至出现发散的问题,为此提出一种基于神经网络的目标跟踪方法,利用深度神经网络解决不同运动模式下目标方位跟踪的问题。水下目标跟踪的神经网路模型可通过运动模型生成大量量测数据进行充分训练,有效解决水声目标数据少、标记样本不足的问题;在量测不连续条件下,提出一种新的损失函数用于增强目标跟踪模型的稳健性;对未学习的仿真数据及实测海试数据进行测试。研究结果表明:构建的卷积神经网络(Convolutional Neural Network, CNN)模型适用于3种不同运动模式下的目标,能在平台静止和运动两种情况下稳定跟踪目标;CNN模型较传统Kalman滤波方法跟踪误差分别降低了7.75°和1.41°,验证了该模型的稳健性和可推广性。

关键词: 水下目标跟踪, 深度学习, 弱目标跟踪, 纯方位跟踪

Abstract:

The tracking error of the traditional Kalman method is significantly increased or even divergent due to the influence of interference and noise. A neural network-based target tracking method is proposed. The proposed method uses the deep neural network to solve the nonlinear change of target azimuth with time in different motion modes. The neural network model of underwater target tracking can generate a large number of measurement data through the motion model for full training, which effectively solves the problems of insufficient underwater acoustic target data and insufficient labeled samples. A new loss function is proposed to enhance the robustness of target tracking model under the condition of measurement discontinuity. The unlearned simulation data and measured sea trial data were tested. The results show that the convolutional neural network (CNN) is applicable for a target in 3 different motion modes,and can stably track targets when the platform is stationary or moving. Compared with the traditional Kalman filtering method, the tracking error of the neural network model is reduced by 7.75° and 1.41°, respectively, for the unlearned simulation data and the measured sea trial data, which verifies the robustness and scalability of the model.

Key words: underwater target tracking, deep learning, weak target tracking, bearings-only tracking

中图分类号:

TB566

杨家铭, 潘悦, 王强, 曹怀刚, 高荪培. 水下弱目标跟踪的深度学习方法研究[J]. 兵工学报, 2024, 45(2): 385-394.

YANG Jiaming, PAN Yue, WANG Qiang, CAO Huaigang, GAO Sunpei. Research on Deep Learning Method of Underwater Weak Target Tracking[J]. Acta Armamentarii, 2024, 45(2): 385-394.

图/表 18

图1 目标运动示意图

Fig.1 Schematic diagram of target motion

图2 目标方位时间历程图

Fig.2 Target bearing-time curve

图3 CNN结构

Fig.3 CNN structure

表1 数据集仿真参数

Table 1 Data set simulation parameters

运动模式	参数	数值
	目标与观测平台距离/km	0.05~10
CV	目标初始方位角/(°)	0~360
	目标速度/(m·s^-1)	0.1~10
	目标速度方向/(°)	0~360
	目标与观测平台距离/km	0.05~10
	目标初始方位角/(°)	0~360
CA	目标速度/(m·s^-1)	0.1~10
	目标速度方向/(°)	0~360
	目标加速度/(m·s^-2)	0.1~2
	目标与观测平台距离/km	0.05~10
	目标初始方位角/(°)	0~360
CV+CT	目标速度/(m·s^-1)	0.1~10
	目标速度方向/(°)	0~360
	目标角速度/((°)·s^-1)	-2~2

图4 3种运动模式下的目标方位轨迹

Fig.4 Target bearings trajectory under three motion modes

表2 量测不确定性参数统计

Table 2 Measurement statistics

事件	概率/%
目标被探测	97.82
出现野值	2.31

表3 CNN参数设置

Table 3 Parameter settings of CNN

网络层	卷积核	节点数	输出维数	参数数量
	512×2		199×512	1536
卷积层	256×2		198×256	262400
	256×2		197×256	131328
	128×2		196×128	65664
最大值池化层	128×2		98×128	0
		200	1×200	2509000
全连接层		1000	1×1000	201000
		200	1×200	200200

表4 全连接神经网络参数设置

Table 4 Parameter settingsof fully-connected neural network

网络层	节点数	输出维数	参数数量
	1000	1×1000	201000
	1000	1×1000	1001000
全连接层	800	1×800	800800
	800	1×800	640800
	600	1×600	480600
	400	1×400	240400
输出层	200	1×200	80200

图5 3种运动模式下的目标跟踪结果

Fig.5 Target tracking results under three motion modes

图6 3种运动模式下的目标跟踪偏差

Fig.6 Target tracking errors under three motion modes

表5 测试数据集跟踪RMSE对比

Table 5 Comparison of tracking RMSEs of test data

损失函数形式	网络结构	RMSE/(°)
MSE损失函数	全连接神经网络	3.12
	CNN	1.44
方位关联损失函数	全连接神经网络	2.07
	CNN	0.72

图7 平台一次转弯机动跟踪结果

Fig.7 Tracking results of a turning maneuver of the platform

图8 平台CV+CA运动模式跟踪结果

Fig.8 Tracking results of CV+CA modes of the platform

图9 平台CV+CT运动模式跟踪结果

Fig.9 Tracking results of CV+CT modes of the platform

表6 平台进行一次机动情况下各方法跟踪RMSE

Table 6 Comparison of tracking RMSEs for a maneuver of the platform

机动模式	跟踪方法	RMSE/(°)
	KF	5.26
一次转弯	全连接神经网络	4.51
	CNN	2.41
	KF	5.88
CV+CA	全连接神经网络	4.42
	CNN	2.66
	KF	7.66
CV+CT	全连接神经网络	2.74
	CNN	1.63

图10 HLAN波束时间历程

Fig.10 HLAN beam-time figure

图11 3种方法对HLAN数据跟踪结果对比

Fig.11 Comparison of tracking results for HLAN data using three different methods

表7 SwellEx-96实际数据跟踪RMSE对比

Table 7 Comparison of tracking RMSEs of SwellEx96 data

跟踪方法	RMSE/(°)
KF方法	2.31
全连接神经网络	1.73
CNN	0.90

参考文献 25

[1]	ZHANG L, LAN J, LI X R. Performance evaluation of joint tracking and classification[J]. IEEE Transactions on Systems, Man, and Cybernetics: Systems, 2021, 51(2): 1149-1163. doi: 10.1109/TSMC.6221021 URL
[2]	LI Q H. Digital sonar design in underwater acoustic[M]. Berlin, Germany: Springer Science and Business Media LLC, 2012.
[3]	王宇杰, 李宇, 鞠东豪, 等. 一种基于水下无人航行器的多目标被动跟踪算法[J]. 电子与信息学报, 2020, 42(8): 2013-2020.
	WANG Y J, LI Y, JU D H, et al. A multi-target passive tracking algorithm based on unmanned underwater vehicle[J]. Journal of Electronics & Information Technology, 2020, 42(8): 2013-2020. (in Chinese)
[4]	马艳, 刘小东. 状态自适应无迹卡尔曼滤波算法及其在水下机动目标跟踪中的应用[J]. 兵工学报, 2019, 40(2): 361-368. doi: 10.3969/j.issn.1000-1093.2019.02.016
	MA Y, LIU X D. State adaptive unscented Kalman filter algorithm and its application in tracking of underwater maneuvering target[J]. Acta Armamentarii, 2019, 40(2): 361-368. (in Chinese) doi: 10.3969/j.issn.1000-1093.2019.02.016
[5]	ANDERSON B D O, MOORE J B, ESLAMI M. Optimal filtering[M]. Englewood Cliffs, NJ, US: Prentice-Hall, 1979.
[6]	GORDON N J, SALMOND D J, SMITH A F M. Novel approach to nonlinear/non-Gaussian Bayesian state estimation[J]. IEE Proceedings F: Rader and Signal Processing, 1993, 140(2): 107-113.
[7]	UHLMANN J K. Algorithm for multiple target tracking[J]. American Science, 1992, 80(2): 128-141. doi: 10.1038/scientificamerican08211915-128supp URL
[8]	SINGER R A, STEIN J J. An optimal tracking filter for processing sensor data of imprecisely determined origin in surveillance system[C]// Proceedings of the 10th IEEE Conference on Decision and Control. Miami Beach, FL, US: IEEE, 1971: 171-175.
[9]	BAR-SHALOM Y. Tracking in a cluttered environment with probabilistic data association[J]. Automatica, 1975, 11(5): 451-460. doi: 10.1016/0005-1098(75)90021-7 URL
[10]	BLACKMAN S S. Multiple hypothesis tracking for multitarget tracking[J]. IEEE Transactions on Aerospace and Electronic Systems, 2004, 19(1): 5-18.
[11]	CHIN L. Application of neural networks in target tracking data fusion[J]. IEEE Transactions on Aerospace and Electronic Systems, 1994, 30(1): 281-287. doi: 10.1109/7.250437 URL
[12]	WINTER M, FAVIER G. A neural network for data association[C]//Proceedings of 1999 IEEE International Conference on Acoustics, Speech, and Signal Processing. Phoenix, AZ, US: IEEE, 1999: 1041-1044.
[13]	CHIN L. Application of neural networks aided target tracking[C]//Proceedings of International Conference on Neural Networks (ICNN’96). Washington, DC, US: IEEE, 1996: 1634-1639.
[14]	邵鹏飞, 王蕾, 王方勇. 基于序贯蒙特卡洛与概率假设密度滤波的主动分布式声纳多目标跟踪[J]. 兵工学报, 2020, 41(5): 941-949. doi: 10.3969/j.issn.1000-1093.2020.05.013
	SHAO P F, WANG L, WANG F Y. Active distributed sonar multi-target tracking based on SMC-PHD filtering[J]. Acta Armamentarii, 2020, 41(5): 941-949. (in Chinese) doi: 10.3969/j.issn.1000-1093.2020.05.013
[15]	YIN W P, KANN K, MO Y, et al. Comparative study of CNN and RNN for natural language processing: arXiv:1702.01923[R/OL]. Ithaca, NY, US: Cornell University, 2017(2017-02-07). https://doi.org/10.48550/arXiv.1702.01923.
[16]	GAO C, YAN J K, ZHOU S H, et al. Long short-term memory-based deep recurrent neural networks for target tracking[J]. Information Sciences, 2019, 502: 279-296. doi: 10.1016/j.ins.2019.06.039 URL
[17]	LIU J X, WANG Z L, XU M. DeepMTT: a deep learning maneuvering target-tracking algorithm based on bidirectional LSTM Network[J]. Information Fusion, 2020, 53(1): 289-304. doi: 10.1016/j.inffus.2019.06.012 URL
[18]	李悦, 马晓川, 刘宇, 等. 非等声速信道下的循环神经网络机动目标跟踪模型[J]. 声学学报, 2021, 46(6):1013-1027.
	LI Y, MA X C, LIU Y, et al. Tracking model of maneuvering target based on recurrent neural network in non-equal sound speed channel[J]. Acta Acustica, 2021, 46(6):1013-1027. (in Chinese)
[19]	RAO S K, LAKSHMI M K, GHOSH A. neural extended/unscented kalman filter for submarine passive target tracking[C]//Proceedings of OCEANS 2022. Chennai, India: IEEE, 2022: 1-7.
[20]	BAIRD Z, MCDONALD M K, RAJAN S, et al. A CNN-LSTM network for augmenting target detection in real maritime wide area surveillance radar data[J]. IEEE Access, 2020, 8: 179281-179294. doi: 10.1109/Access.6287639 URL
[21]	SAINATH T N, VINYALS O, SENIOR A, et al. Convolutional long short-term memory fully connected deep neural networks[C]//Proceedings of 2015 IEEE International Conference on Acoustics, Speech and Signal Processing. South Brisbane, QLD, Australia: IEEE, 2015: 4580-4584.
[22]	张永梅, 赖裕平, 马健喆, 等. 基于视频的装甲车和飞机检测跟踪及轨迹预测算法[J]. 兵工学报, 2021, 42(3): 545-554. doi: 10.3969/j.issn.1000-1093.2021.03.010
	ZHANG Y M, LAI Y P, MA J Z, et al. A video-based prediction algorithm for armored vehicle and aricraft detection/tracking and trajectory[J]. Acta Armamentarii, 2021, 42(3): 545-554. (in Chinese)
[23]	DUCHI J, HAZAN E, SINGER Y. Adaptive subgradient methods for online learning and stochastic optimization[J]. Journal of Machine Learing Research, 2011, 12: 2121-2159.
[24]	VO B N, MA W K. The gaussian mixture probability hypothesis density filter[J]. IEEE Transactions on Signal Processing, 2006, 54(11): 4091-4104. doi: 10.1109/TSP.2006.881190 URL
[25]	MURRAY J, ENSBERG D. The SwellEx-96 experiment[EB/OL]. (2003-04-29)[2022-09-16]. http://swellex96.ucsd.edu/events.htm.