Correlation Filter Object Tracking Based on Adaptive Spatiotemporal Regularization

Xiangming Qi; Wei Chen

doi:10.3788/LOP202158.0215007

Journals >Laser & Optoelectronics Progress >Volume 58 >Issue 2 >Page 0215007 > Article

Laser & Optoelectronics Progress
Vol. 58, Issue 2, 0215007 (2021)

Correlation Filter Object Tracking Based on Adaptive Spatiotemporal Regularization

Xiangming Qi and Wei Chen^*

Author Affiliations

College of Software, Liaoning Technical University, Huludao, Liaoning 125105, China

show less

DOI: 10.3788/LOP202158.0215007 Cite this Article Set citation alerts

Xiangming Qi, Wei Chen. Correlation Filter Object Tracking Based on Adaptive Spatiotemporal Regularization[J]. Laser & Optoelectronics Progress, 2021, 58(2): 0215007 Copy Citation Text

show less

Process of obtaining initial frame saliency awareness reference weight for Soccer video sequence. (a) First frame image; (b) saliency detection result; (c) result of saliency detection after treatment; (d) saliency image; (e) saliency awareness reference weight for initial frame

Fig. 1. Process of obtaining initial frame saliency awareness reference weight for Soccer video sequence. (a) First frame image; (b) saliency detection result; (c) result of saliency detection after treatment; (d) saliency image; (e) saliency awareness reference weight for initial frame

Download full size

Visualization of four different spatial regularization weights. (a) W1; (b) W2; (c) W3; (d) W4

Fig. 2. Visualization of four different spatial regularization weights. (a) W₁; (b) W₂; (c) W₃; (d) W₄

Download full size

Fig. 3. Response of object in different cases for Box video sequence. (a) Object for 130^th frame; (b) response for 130^th frame; (c) object for 460^th frame; (d) response for 460^th frame

Download full size

Fig. 4. Distance precision and success rate on OTB-2015 dataset. (a) Distance precision curve; (b) success rate curve

Download full size

Fig. 5. Success rate of 4 different attribute video sequences on OTB-2015 dataset. (a) Deformation; (b) out-of-plane rotation; (c) occlusion; (d) out of view

Download full size

Fig. 6. Tracking results of our algorithm and comparison algorithms on 4 video sequences. (a) Box; (b) dragonbaby; (c) shaking; (d) soccer

Download full size

ADMM	Relevant parameter
Solving filtering function and temporal regularization parameter	γ⁽⁰⁾=1,γ^max=10,α=0.1,N_I=2
Solving spatial regularization weight	ζ⁽⁰⁾=1,ζ^max=100,β=10,N'_I=2

Table 1. Relevant parameters setting of ADMM in the process of algorithm optimization

Parameter setting	DP	SR
ζ⁽⁰⁾=1,ζ^max=1000,β=10,N'_I=2	86.2	65.1
ζ⁽⁰⁾=1,ζ^max=10000,β=10,N'_I=2	84.1	63.8
ζ⁽⁰⁾=1,ζ^max=100,β=10,N'_I=2	86.4	65.6
ζ⁽⁰⁾=1,ζ^max=100,β=10,N'_I=3	84.9	64.3

Table 2. Comparison of tracking performance of algorithm in different parameter settingsunit: %

Algorithm	DSST	BACF	SRDCF	STRCF	CNN-SVM	Ours
Average tracking speed /(frame·s^-1)	31.8	22.1	3.3	13.7	N	12.1

Table 3. Comparison of average tracking speeds on OTB-2015 dataset

Xiangming Qi, Wei Chen. Correlation Filter Object Tracking Based on Adaptive Spatiotemporal Regularization[J]. Laser & Optoelectronics Progress, 2021, 58(2): 0215007

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information