# 基于核相关滤波和卡尔曼滤波预测的混合跟踪方法

(郑州大学 电气与信息工程学院,河南 郑州 450001)

## 1 核相关滤波算法原理

KCF 算法以岭回归方式训练滤波器,假设训练样本为xi,其目的是寻找一个目标分类器f(xi)=wTxi,使得滤波器期望输出yi和训练样本xi之间的均方误差函数最小,即

(1)

w=(XTX+λI)-1XTy

(2)

(3)

α=(K+λI)-1y

(4)

(5)

(6)

KCF目标跟踪算法采用线性插值并且以固定的学习率对模型进行更新,更新方式如下:

(7)

## 2 改进的核相关滤波算法

Figure 1 Process of target tracking algorithm based on improved kernel correlation filtering

(8)

(9)

### 2.3 卡尔曼滤波预测跟踪算法

KCF目标跟踪算法基于目标在上一帧中所处位置的特征相似性来搜索目标在当前帧中的位置,没有利用目标的运动信息,并过分依赖特征最大响应值来确定目标位置[13]。而KF对运动目标具有良好的运动约束,可以有效地减少跟踪失败[13]。因此,基于KCF跟踪框架,可以将KF视为运动信息的补充,KF通过对目标状态的测量和预测来估计其当前状态。跟踪目标的运动状态方程为

xk=Fkxk-1+wk;

(10)

yk=Hkxk+vk

(11)

(12)

(13)

KF算法可分为2个步骤:预测步骤和更新步骤。

(14)

(15)

(16)

(17)

Pk|k=(I-KkHk)Pk|k-1

(18)

## 3 实验结果与分析

### 3.2 跟踪性能分析

Table 1 Performance metrics of five tracking algorithms

Figure 2 Accuracy curve and success rate curve of different tracking algorithms

Figure 3 Tracking results of different algorithms on occluded video sequences

## 4 结论

[1] 孟晓燕, 段建民. 基于相关滤波的目标跟踪算法研究综述[J]. 北京工业大学学报, 2020, 46(12): 1393-1416.

MENG X Y, DUAN J M. Advances in correlation filter-based object tracking algorithms:a review[J].Journal of Beijing University of Technology,2020,46(12):1393-1416.

[2] BOLME D S, BEVERIDGE J R, DRAPER B A, et al. Visual object tracking using adaptive correlation filters[C]∥2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE,2010:2544-2550.

[3] HENRIQUES J F, CASEIRO R, MARTINS P, et al. Exploiting the circulant structure of tracking-by-detection with kernels[C]∥Proceedings of the 12th European Conference on Computer Vision. New York: ACM, 2012: 702-715.

[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] DANELLJAN M, HGER G, KHAN F S, et al. Discriminative scale space tracking[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, 39(8): 1561-1575.

[6] LI Y, ZHU J K. A scale adaptive kernel correlation filter tracker with feature integration[C]∥Computer Vision-ECCV 2014 Workshops. Cham: Springer, 2015: 254-265.

[7] BERTINETTO L, VALMADRE J, GOLODETZ S, et al. Staple: complementary learners for real-time tracking[C]∥2016 IEEE Conference on Computer Vision and Pattern Recognition. Piscataway:IEEE,2016:1401-1409.

[8] KALAL Z, MIKOLAJCZYK K, MATAS J. Tracking-learning-detection[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2012, 34(7): 1409-1422.

[9] 陈丽萍, 王铭羽, 杨文柱, 等. 基于改进核相关滤波的长时目标跟踪算法[J]. 郑州大学学报(工学版), 2020, 41(3): 32-36.

CHEN L P, WANG M Y, YANG W Z, et al. An improved kernelized correlation filter for long-term target tracking[J]. Journal of Zhengzhou University (Engineering Science), 2020, 41(3): 32-36.

[10] 潘振福, 朱永利. 使用PSR重检测改进的核相关目标跟踪方法[J]. 计算机工程与应用, 2017, 53(12): 196-202.

PAN Z F, ZHU Y L. Improvement for tracker with kernelized correlation filters with PSR redetection[J]. Computer Engineering and Applications, 2017, 53(12): 196-202.

[11] CHEN H, ZHANG W G, YAN D H. Robust visual tracking with reliable object information and Kalman filter[J]. Sensors (Basel), 2021, 21(3): 889.

[12] YUAN Y, CHU J, LENG L, et al. A scale-adaptive object-tracking algorithm with occlusion detection[J]. EURASIP Journal on Image and Video Processing, 2020, 2020(1): 7.

[13] LI F Q, DONG E Z, YANG S. An improved object tracking algorithm combining LCT kernel correlation filtering and Kalman prediction[C]∥2021 China Automation Congress (CAC). Piscataway: IEEE, 2021:877-882.

[14] 刘耀胜,廖育荣,林存宝,等.基于特征融合与抗遮挡的卫星视频目标跟踪方法[J].北京航空航天大学学报,2022,48(12):2537-2547.

LIU Y S,LIAO Y R,LIN C B, et al. Feature-fusion and anti-occlusion based target tracking method for satellite videos[J].Journal of Beijing University of Aeronautics and Astronautics,2022,48(12):2537-2547.

[15] WU Y, LIM J, YANG M H. Online object tracking: a benchmark[C]∥2013 IEEE Conference on Computer Vision and Pattern Recognition. Piscataway:IEEE, 2013:2411-2418.

# Hybrid Tracking Method Based on Kernel Correlation Filter and Kalman Filter Prediction

FAN Wenbing, ZHANG Lulu

(School of Electrical and Information Engineering, Zhengzhou University, Zhengzhou 450001, China)

AbstractAiming at the problem that KCF tracking algorithm might decrease the tracking performance or even tracks unsuccessfully in the occlusion scene, an anti-occlusion model adaptation image tracking algorithm was proposed by combining KCF and KF prediction. Firstly, considering the lack of occlusion evaluation in the traditional KCF target tracking algorithm, the peak sidelobe rate of the response map was introduced to judge the occlusion of the image target, and the occlusion types were divided into partial occlusion and severe occlusion. Then different model update strategies were adopted according to the severity of occlusion. When the target was not occluded or occluded partially, instead of using a fixed learning rate to update the model in the traditional KCF tracking algorithm, the target appearance model was updated by adjusting the model learning rate adaptively to avoid tracking drift. When the target was severely occluded, stopped updating the KCF model. Finally, the state space and position output models of Kalman filter were constructed by applying the motion information before severe occlusion. The Kalman filter prediction algorithm was designed to predict the moving target trajectory and estimate the target position in the occlusion scene,so as to solve the problem of target tracking failure in occlusion scenes. The OTB-2013 standard dataset was utilized to conduct extensive experiments, the results demonstrated that the distance accuracy of the proposed hybrid tracking algorithm KCF-KF was 0.796, and the overlap success rate was 0.692. Compared with the other traditional tracking algorithms, the tracking accuracy and success rate of the hybrid algorithm were better, and the hybrid algorithm could achieve better tracking performance when encountering the target occlusion challenges and solve the occlusion interference in the tracking process effectively.

Keywordskernel correlation filter; occlusion; peak sidelobe rate; adaptive model updating; Kalman filter

doi：10.13705/j.issn.1671-6833.2023.02.014