改进的核相关滤波跟踪算法

doi:10.16180/j.cnki.issn1007-7820.2020.03.010

摘要/Abstract

摘要：

针对核相关滤波算法在目标跟踪过程中尺度特定和遮挡判断失败的问题,文中提出一种利用自适应特征融合的位置滤波器来判断目标是否被遮挡的方法。该方法检测到峰值旁瓣比异常时,停止模型自适应更新,启动在线重检测;并结合尺度金字塔中的尺度滤波器来确定目标尺寸,从而得出精准的目标位置。实验通过复杂背景下的10组运动视频来评估改进算法的性能。与基础核相关滤波算法相比,改进算法的平均中心位置误差降低了36.683 pixel;在像素阈值设为20 pixel时,平均距离精度提升了44.632%;在边界框重叠阈值设为0.5时,重叠精度提升了46.453%。

关键词: 目标跟踪, 特征融合, 遮挡判别, 目标模型更新, 尺度滤波器, 位置滤波器

Abstract:

In order to solved the probelmof scale specificity and occlusion judgment failure of Kernel-correlation Filtering algorithm in target tracking, a position filter based on adaptive feature fusion was proposed to judge whether the target was occluded or not. When the Peak-to- Sidelobe Ratio anomaly was detected, the adaptive updating of the model was stopped and online re-detection was started, and the target size was determined by combining the scale filter in the scale pyramid, thus the accurate target location was obtained. The experiment evaluated the performance of the improved algorithm through 10 groups of motion video in complex background. Compared with the basic Kernel-correlation Filtering algorithm, the average center position error of the improved algorithm was reduced by 36.683 pixels; the average distance accuracy was increased by 44.632% when the threshold of the pixel was set to 20 pixels; and the overlap accuracy was increased by 46.453% when the boundary frame overlap threshold was set to 0.5.

Key words: target tracking, featurefusion, occlusiondiscrimination, modelupdate, scalefilter, translation filter

中图分类号:

TP391.41

曾照,吴薇,汪欣. 改进的核相关滤波跟踪算法[J]. 电子科技, 2020, 33(3): 50-55.

ZENG Zhao,WU Wei,WANG Xin. Improved Kernelized Correlation Filter Tracking[J]. Electronic Science and Technology, 2020, 33(3): 50-55.

图/表 11

图1

图2

图3

图4

图5

表1

表2

表3

表4

图6

图7

参考文献 16

[1]	尹宏鹏, 陈波, 柴毅 , 等. 基于视觉的目标检测与跟踪综述[J]. 自动化学报, 2016,42(10):1466-1489.
	Yin Hongpeng, Chen Bo, Chai Yi , et al. Asurvey of vision-based target detection and tracking[J]. Journal of Automation, 2016,42(10):1466-1489.
[2]	刘紫燕, 唐虎, 冯亮 . 基于FPGA的机器视觉技术视频跟踪系统实现[J]. 电子科技, 2018,31(9):41-44,49.
	Liu Ziyan, Tang Hu, Feng Liang . Implementation of machine vision technology video tracking system based on FPGA[J]. Electronic Science and Technology, 2018,31(9):41-44,49.
[3]	Chen Z, Hong Z, Tao D . An experimental survey on correlation filter-based tracking[J]. Computer Science, 2015,53(25):68-83.
[4]	Henriques J F, Caseiro R, Martins P , et al. High-speed tracking with kernelized correlation filters[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2015,37(3):583-596.
[5]	张雷, 于凤芹 . 基于置信图特性的改进时空上下文目标跟踪[J]. 计算机工程, 2016,42(8):277-281,288.
	Zhang Lei, Yu Fengqin . Improved spatiotemporal context target tracking based on confidence graph characteristics[J]. Computer Engineering, 2016,42(8):277-281,288.
[6]	Danelljan M, Khan F S, Felsberg M, et al. Adaptive color attributes for real-time visual tracking[C]. Columbus: IEEE Conference on Computer Vision and Pattern Recognition, 2014.
[7]	张雷, 王延杰, 孙宏海 , 等. 采用核相关滤波器的自适应尺度目标跟踪[J]. 光学精密工程, 2016,24(2):448-459.
	Zhang Lei, Wang Yanjie, Sun Honghai , et al. Adaptivescale target tracking using kernel correlation filter[J]. Optical and Precision Engineering, 2016,24(2):448-459.
[8]	Danelljan M, Hger G, Khan F S, et al. Accurate scale estimation for robust visual tracking[C]. British: British Machine Vision Conference, 2014.
[9]	Cheon M, Lee W, Yoon C , et al. Vision-based vehicle detection system with consideration of the detecting location[J]. IEEE Transactions on Intelligent Transportation Systems, 2012,13(3):1243-1252.
[10]	李远状, 韩彦芳, 于书盼 . 一种核相关滤波器的多尺度目标跟踪方法[J]. 电子科技, 2018,31(10):1-5,10.
	Li Yuanzhuang, Han Yanfang, Yu Shupan . Amulti-scale target tracking method based on core-correlation filter[J]. Electronic Science and Technology, 2018,31(10):1-5,10.
[11]	Possegger H, Mauthner T, Bischof H. In defense of color-based model-free tracking[C]. Boston:IEEE Conference on Computer Vision and Pattern Recognition, 2015.
[12]	Ma C, Yang X, Zhang C, et al. Long-term correlation tracking[C]. Boston:IEEE Conference on Computer Vision & Pattern Recognition, 2015.
[13]	Wang M, Liu Y, Huang Z. Large margin object tracking with circulant feature maps[C]. Honolulu:IEEE Conference on Computer Vision and Pattern Recognition, 2017.
[14]	Li Y, Zhu J. A scale adaptive kernel correlation filter tracker with feature integration[C]. Zurich:European Conference on Computer Vision, 2014.
[15]	Yi W, Lim J, Yang M H. Online object tracking: a benchmark[C]. Portland:IEEE Conference on Computer Vision & Pattern Recognition, 2013.
[16]	Joao F Henriques, Caseiro R, Martins P, et al. Exploiting the circulant structure of tracking-by-detection with kernels[C]. Florence: European Conference on Computer Vision, 2012.

视频	帧数	主要影响因素
Jogging1	307	尺度、遮挡、快速运动、旋转
Bird2	99	背景干扰、遮挡、快速运动、旋转
Jumping	313	快速运动、模糊、遮挡
Freeman4	297	快速运动、背景干扰、尺度、遮挡、
Shaking	365	光照、遮挡、快速运动、背景干扰
Surfer	376	快速运动、模糊、背景干扰、旋转
ClifBar	472	背景干扰、尺度、快速运动、旋转
BlurBody	334	模糊、遮挡、形变、旋转、背景干扰
Couple	140	光照、快速运动、背景干扰
Woman	597	光照、尺度、遮挡、模糊、快速运动

视频	Ours	KCF	DSST	CN
Jogging1	4.22	87.90	110.00	101.00
Bird2	15.70	21.40	55.70	5.99
Jumping	4.44	26.20	37.50	61.60
Freeman4	10.10	27.10	5.43	46.00
Shaking	6.92	113.00	8.08	15.10
Surfer	5.39	8.51	20.10	9.93
ClifBar	8.31	37.10	5.45	31.60
BlurBody	8.71	64.20	91.20	38.60
Couple	11.40	47.20	125.00	123.00
Woman	9.40	10.10	9.67	283.00
平均	7.588	44.271	46.813	71.582

视频	Ours	KCF	DSST	CN
Jogging1	97.068	23.453	23.127	23.799
Bird2	77.778	47.475	47.475	98.990
Jumping	98.083	33.866	5.112	5.112
Freeman4	87.633	51.943	95.760	26.148
Shaking	98.356	1.918	100.000	69.863
Surfer	97.872	90.957	69.149	79.255
ClifBar	93.856	44.280	99.788	60.169
BlurBody	97.006	58.383	61.976	54.192
Couple	76.492	25.714	10.714	10.714
Woman	93.970	93.802	93.802	24.958
平均	91.811	47.179	60.690	45.32

视频	Ours	KCF	DSST	CN
Jogging1	78.2	22.50	22.50	22.50
Bird2	76.8	47.50	47.50	99.00
Jumping	93.6	28.10	4.790	4.79
Freeman4	54.4	18.40	45.90	17.30
Shaking	98.6	1.37	100.00	67.10
Surfer	93.9	38.60	29.00	34.80
ClifBar	69.5	29.90	88.30	27.60
BlurBody	99.4	58.70	62.30	80.80
Couple	66.4	24.30	10.70	10.70
Woman	96.7	93.60	93.30	24.30
平均	82.75	36.297	50.429	38.889