通道裁剪下的多特征组合目标跟踪算法

doi:10.3969/j.issn.1001-506X.2020.04.05

摘要/Abstract

摘要：

针对深层特征存在冗余通道影响跟踪速度和精度以及单一特征难以适应复杂场景的问题,提出了一种通道裁剪下的多特征组合目标跟踪算法。首先,在相关滤波算法的框架上结合传统手工特征和深层特征进行跟踪。其次,通过对比深层特征通道上目标区域和搜索区域的特征均值设计通道裁剪策略,筛选出合适的特征通道。最后,通过隔帧更新的方式更新深度特征,通过平均峰值相关能量更新传统特征滤波模板,最终实现准确跟踪。与10种算法在OTB2013和OTB2015数据集上进行对比实验的结果表明,本文算法在跟踪准确度和成功率方面都取得了更为理想的结果。

关键词: 目标跟踪, 通道裁剪, 相关滤波, 特征组合

Abstract:

Aiming to solve the problems that the redundant channel in depth features affects the tracking speed and accuracy, and the single feature cannot adapt to all complex scenes, a multi-feature combined target tracking algorithm based on channel clipping is proposed. Firstly, the traditional manual features and depth features are combined for tracking. Secondly, a channel clipping strategy is designed by comparing the deep feature means of each channel in the target and searching areas. Then, the depth feature is updated by the proposed interval updating strategy, and the traditional feature filter template is updated by the average peak correlation energy. Compared with the 10 algorithms on the OTB2013 and OTB2015 datasets, the results show that the proposed algorithm achieves better results in tracking accuracy and success rate.

Key words: object tracking, channel clipping, correlation filter, feature combination

中图分类号:

TP391

谢瑜, 陈莹. 通道裁剪下的多特征组合目标跟踪算法[J]. 系统工程与电子技术, 2020, 42(4): 764-772.

Yu XIE, Ying CHEN. Multi-feature combined target tracking algorithm based on channel clipping[J]. Systems Engineering and Electronics, 2020, 42(4): 764-772.

图/表 14

图1

图2

图3

图4

图5

表1

图6

图7

表2

表3

图8

图9

图10

表4

参考文献 14

1	MA C , HUANG J B , YANG X , et al. Adaptive correlation filters with long-term and short-term memory for object tracking[J]. International Journal of Computer Vision, 2018, 126 (8): 771- 796. doi: 10.1007/s11263-018-1076-4
2	HENRIQUES J F , CASEIRO R , MARTINS P , et al. High-speed tracking with kernelized correlation filters[J]. IEEE Trans.on Pattern Analysis and Machine Intelligence, 2015, 37 (3): 853- 896.
3	BERTINETTO L, VALMADRE J, GOLODETZ S, et al.Staple: complementary learners for real-time tracking[C]//Proc.of the IEEE Conference on Computer Vision and Pattern Recognition, 2016: 1401-1409.
4	MA C, HUANG J B, YANG X, et al.Hierarchical convolutional features for visual tracking[C]//Proc.of the IEEE International Conference on Computer Vision, 2015: 3074-3082.
5	WANG M, LIU Y, HUANG Z.Large margin object tracking with circulant feature maps[C]//Proc.of the IEEE Conference on Computer Vision and Pattern Recognition, 2017: 4021-4029.
6	LI X, MA C, WU B Y, et al.Target-aware deep tracking[C]//Proc.of the IEEE Conference on Computer Vision and Pattern Recognition, 2019: 1-10.
7	ZHANG T, XU C, YANG M H.Multi-task correlation particle filter for robust object tracking[C]//Proc.of the IEEE Conference on Computer Vision and Pattern Recognition, 2017: 4819-4827.
8	YOO S, YUN K, CHOI J Y, et al.Action-decision networks for visual tracking with deep reinforcement learning[C]//Proc.of the IEEE Conference on Computer Vision and Pattern Recognition, 2017: 1349-1358.
9	QI Y, ZHANG S, QIN L, et al.Hedged deep tracking[C]//Proc.of the IEEE Conference on Computer Vision and Pattern Recognition, 2016: 4303-4311.
10	MUELLER M, SMITH N, GHANEM B.Context-aware correlation filter tracking[C]//Proc.of the IEEE Conference on Computer Vision and Pattern Recognition, 2017: 1396-1404.
11	MA C , HUANG J B , YANG X , et al. Robust visual tracking via hierarchical convolutional features[J]. IEEE Trans.on Pattern Analysis and Machine Intelligence, 2018.
12	MA C, YANG X, ZHANG C, et al.Long-term correlation tracking[C]//Proc.of the IEEE Conference on Computer Vision and Pattern Recognition, 2015: 5388-5396.
13	WU Y, LIM J, YANG M H.Online object tracking: a benchmark[C]//Proc.of the IEEE Conference on Computer Vision and Pattern Recognition, 2013: 2411-2418.
14	WU Y , LIM J , YANG M H . Object tracking benchmark[J]. IEEE Trans.on Pattern Analysis and Machine Intelligence, 2015, 37 (9): 1834- 1848. doi: 10.1109/TPAMI.2014.2388226

算法名称	OTB2013			OTB2015			fps
算法名称	DP/%	OS/%	CLE/像素	DP/%	OS/%	CLE/像素	fps
FCCT	87.5	79.7	16.7	90.8	85.3	12.2	14.5
TADT^[10]	83.6	76.9	18.2	86.4	81.3	16.3	33.7
ADNet^[13]	83.4	74.6	20.9	87.9	80.8	14.8	2.9
LSCT^[12]	76.8	66.5	34.3	83.0	74.6	24.0	13.8
HCFTs^[16]	84.2	68.9	22.8	87.5	72.7	16.2	0.7
MCPF^[11]	84.9	72.0	24.9	87.5	78.3	20.7	0.5
HDT^[14]	80.9	60.2	24.7	84.9	66.3	20.0	5.5
CACF^[15]	73.4	64.9	41.8	80.8	72.8	31.8	38.5
HCF^[4]	81.0	69.7	27.8	83.1	74.2	24.8	10.3
LCT^[17]	70.1	63.5	103.2	76.5	70.6	66.5	20.7
KCF^[6]	61.5	47.3	53.6	69.5	55.6	44.8	217.7

参数	FCCT	ADNet	TADT	HCFTs	LSCT	MCPF	HDT	HCF	CACF	KCF	LCT
光照变化	91.1	90.3	86.1	90.2	84.3	88.6	82.3	84.3	82.5	71.8	75.4
平面外旋转	90.7	87.8	87.2	84.9	82.6	86.7	80.5	80.8	76.4	67.7	74.6
尺度变化	88.3	86.6	86.3	83.6	79.1	86.7	81.3	80.1	76.2	63.7	69.0
遮挡	85.9	83.9	84.5	81.7	76.7	86.5	77.8	77.8	74.5	63.7	68.8
形变	87.2	84.8	81.4	81.8	77.4	80.7	81.3	75.8	76.7	59.9	68.0
运动模糊	88.4	85.3	83.9	82.8	79.0	84.5	79.6	80.7	75.7	61.4	68.0
平面内旋转	92.0	87.1	83.5	89.7	85.1	89.0	84.7	86.8	80.0	70.7	78.1
出视野	83.1	80.2	81.6	74.6	67.1	76.4	66.3	66.2	66.6	50.1	59.2
背景混杂	91.4	90.9	79.9	90.3	84.8	82.7	84.8	82.4	79.2	71.2	74.4
快速运动	87.7	83.3	81.9	80.9	79.5	83.0	80.4	79.4	74.7	61.7	67.3
低分辨率	82.7	85.1	87.7	88.1	77.0	94.1	77.2	83.2	56.3	52.4	49.6

参数	FCCT	ADNet	TADT	HCFTs	LSCT	MCPF	HDT	HCF	CACF	KCF	LCT
光照变化	87.6	84.5	82.6	74.7	76.9	76.7	62.3	78.9	73.8	56.2	73.1
平面外旋转	84.0	77.7	80.6	68.5	72.5	76.7	62.7	70.6	67.2	52.7	67.6
尺度变化	80.2	78.8	79.9	60.7	66.9	73.8	52.9	67.2	64.8	43.1	59.9
遮挡	81.8	75.7	81.0	67.3	71.6	75.8	61.9	71.2	69.1	52.2	64.1
形变	79.8	75.8	74.0	66.6	69.4	69.5	62.1	66.6	69.4	50.0	61.8
运动模糊	82.9	83.8	83.3	75.0	73.6	72.5	70.0	76.5	71.7	56.5	67.0
平面内旋转	81.7	76.1	76.0	71.2	70.6	77.0	64.7	71.7	68.9	54.5	68.7
出视野	70.5	72.3	74.5	61.8	61.9	62.9	54.7	58.4	57.6	45.7	53.1
背景混杂	86.4	84.5	76.6	80.4	76.6	76.1	73.5	76.8	72.3	62.7	72.0
快速运动	81.5	78.5	78.2	69.2	72.1	70.8	65.8	73.1	69.3	52.7	64.9
低分辨率	68.5	74.1	79.5	64.0	61.6	72.3	46.1	65.2	45.9	32.9	42.3

通道衡量条件	具体含义
feature_cut_1	中位数的时域均值
feature_cut_2	中位数的时域中位数
feature_cut_3	中位数的时域最大值
feature_cut_4	中位数的时域最小值
feature_cut_5	阈值1
feature_cut_6	阈值1.5

[1]	仇祝令, 查宇飞, 李振宇, 李禹铭, 张鹏, 朱川. 基于多模型蒸馏的时间正则化相关滤波跟踪算法[J]. 系统工程与电子技术, 2022, 44(8): 2448-2456.
[2]	侯子林, 程婷, 彭瀚. 基于量测转换序贯滤波的GMPHD机动目标跟踪[J]. 系统工程与电子技术, 2022, 44(8): 2474-2482.
[3]	史浩然, 卢发兴, 祁江鑫, 杨光. 基于辅助信标的无人机协同目标跟踪[J]. 系统工程与电子技术, 2022, 44(7): 2302-2310.
[4]	金国栋, 薛远亮, 谭力宁, 许剑锟. 基于孪生神经网络的目标跟踪算法进展研究[J]. 系统工程与电子技术, 2022, 44(6): 1805-1822.
[5]	翟光, 王妍欣, 孙一勇. 基于低轨星网的多目标协同跟踪滤波技术[J]. 系统工程与电子技术, 2022, 44(6): 1957-1967.
[6]	王帅, 向建军, 彭芳, 唐书娟. 基于新最速下降法的目标跟踪算法[J]. 系统工程与电子技术, 2022, 44(5): 1512-1519.
[7]	辛怀声, 曹晨. 基于交互多模型的分组δ-广义标签多伯努利算法[J]. 系统工程与电子技术, 2022, 44(4): 1128-1138.
[8]	谢家豪, 黄树彩, 韦道知, 张曌宇, 王文豪. 基于P_EV准则的不确定混合多传感器联盟求解[J]. 系统工程与电子技术, 2022, 44(3): 819-826.
[9]	宋子壮, 杨嘉伟, 张东方, 王诗强, 张硕. 基于无锚框的红外多类别多目标实时跟踪网络[J]. 系统工程与电子技术, 2022, 44(2): 401-409.
[10]	姚云翔, 陈莹. 注意力机制下双模态交互融合的目标跟踪网络[J]. 系统工程与电子技术, 2022, 44(2): 410-419.
[11]	辛怀声, 宋鹏汉, 曹晨. 多模型广义标签多伯努利滤波器[J]. 系统工程与电子技术, 2022, 44(12): 3603-3613.
[12]	刘浩楠, 宋骊平. 基于核Fisher判别的群结构更新模型及群目标跟踪算法[J]. 系统工程与电子技术, 2022, 44(10): 3012-3019.
[13]	方澄, 路稳, 姬菁颖, 宋玉蒙, 梁斐菲, 罗志伟. 基于外观相似性更新的相关滤波跟踪算法[J]. 系统工程与电子技术, 2022, 44(1): 117-126.
[14]	黄月平, 李小锋, 杨小冈, 齐乃新, 卢瑞涛, 张胜修. 基于相关滤波的视觉目标跟踪算法新进展[J]. 系统工程与电子技术, 2021, 43(8): 2051-2065.
[15]	赵跃新, 齐望东, 刘鹏, 袁恩, 徐兵. 三维AoA目标跟踪的二次约束卡尔曼滤波算法[J]. 系统工程与电子技术, 2021, 43(8): 2263-2272.