Correlation-Filter Tracking Algorithm with Adaptive-Feature Fusion and Anti-Occlusion

Haifeng Liu; Cheng Sun; Xingliang Liang

doi:10.3788/LOP57.141014

Journals >Laser & Optoelectronics Progress >Volume 57 >Issue 14 >Page 141014 > Article

Laser & Optoelectronics Progress
Vol. 57, Issue 14, 141014 (2020)

Correlation-Filter Tracking Algorithm with Adaptive-Feature Fusion and Anti-Occlusion

Haifeng Liu^1、2, Cheng Sun^1、*, and Xingliang Liang²

Author Affiliations

¹School of Electronic Information and Artificial Intelligence, Shaanxi University of Science & Technology, Xi'an, Shaanxi 710021, China;

²School of Arts and Sciences, Shaanxi University of Science & Technology, Xi'an, Shaanxi 710021, China;

show less

DOI: 10.3788/LOP57.141014 Cite this Article Set citation alerts

Haifeng Liu, Cheng Sun, Xingliang Liang. Correlation-Filter Tracking Algorithm with Adaptive-Feature Fusion and Anti-Occlusion[J]. Laser & Optoelectronics Progress, 2020, 57(14): 141014 Copy Citation Text

show less

Comparison results of tracking confidence of partial frames of Girl2 video sequence under different occlusions. (a) Not occlusion; (b) slight occlusion; (c) severe occlusion

Fig. 1. Comparison results of tracking confidence of partial frames of Girl2 video sequence under different occlusions. (a) Not occlusion; (b) slight occlusion; (c) severe occlusion

Download full size

Fig. 2. Framework of proposed algorithm

Download full size

Fig. 3. Effect of fixed fusion factor on tracking. (a) Precision plot of Bird1; (b) CLE plot of Bird1; (c) precision plot of Bird2; (d) CLE plot of Bird2; (e) precision plot of Board; (f) CLE plot of Board; (g) precision plot of Coke; (h) CLE plot of Coke

Download full size

Fig. 4. Tracking results of Twinnings sequence with different fusion weights coefficients. (a) Tracking results of 195 frame with w_hist=0; (b) tracking results of 250 frame with w_hist=0; (c) tracking results of 195 frame with w_hist=0.1; (d) tracking results of 250 frame with w_hist=0.1; (e) tracking results of 195 frame with w_hist=0.3; (f) tracking results of 250 frame with w_hist=0.3

Download full size

Fig. 5. Test results of experiment 3 on OTB-2013 dataset. (a) OPE precision plot; (b) success rate plot

Download full size

Fig. 6. Test results of experiment 3 on OTB-100 dataset. (a) OPE precision plot; (b) success rate plot

Download full size

Fig. 7. Comparison of tracking results of proposed algorithm with other 4 algorithms on different video sequences. (a) Joggong-2; (b) Coke; (c) Matrix; (d) Football1; (e) Freeman4

Download full size

Algorithm 1: Update model
Require: R_t, M_t_{, MSE}, g_max of response g_t
Ensure: Update model
Occlusion condition:
Not occlusion: R_t>R_threshold1 && g_max>g_threshold11) Update learning rate η_t_{, temp}, model weights w_t_{, temp} and w_t_{, hist}2) Update models h_t from h_t_{, new} and h_t_-1 like formula (9)3) Update color histogram ρ_t from ρ_t_{, new} and ρ_t_-1 like formula (9)
Severe occlusion: R_t<R_threshold2 && g_max<g_threshold2Do not update models h_t=h_t_-1 and color histogram ρ_t=ρ_t_-1Slight occlusion: otherwiseSimilar to Not occlusion, but update rate η_t_{, temp} is smaller in this part

Table 1. Model update strategy

Algorithm	Channelreliability	Adaptivemerge	Updatescheme	OTB-2013		OTB-100
Algorithm	Channelreliability	Adaptivemerge	Updatescheme	Precision /%	Success /%	Precision /%	Success /%
Staple_CR	√			82.1	61.3	79.7	59.5
Staple_AM		√		81.3	61.4	79.2	58.9
Staple_Update			√	79.3	60.8	78.9	58.4
Staple				78.8	59.3	78.4	57.9

Table 2. Setting of tracking model in experiment 2 and OPE test results of each model on OTB

Attribute	SV	IV	OPR	OCC	BC	DEF	MB	FM	IPR	OV	LR
Ours	$\bar{56.3}$	$\bar{58.8}$	$\underset{=}{61.1}$	$\underset{=}{61.5}$	$\bar{58.5}$	$\underset{=}{65.4}$	$\bar{58.0}$	$\underset{=}{58.2}$	$\underset{=}{59.4}$	$\underset{=}{61.1}$	30.3
fDSST	$\underset{=}{57.1}$	$\underset{=}{59.7}$	$\bar{57.2}$	55.8	$\underset{=}{61.7}$	56.4	$\underset{=}{59.3}$	$\bar{55.6}$	$\bar{58.4}$	$\bar{55.5}$	$\bar{38.9}$
DSST	44.7	50.4	49.2	47.8	49.8	47.8	42.3	40.5	52.0	46.5	32.5
Staple	54.5	56.1	56.9	$\bar{58.5}$	55.7	$\bar{60.7}$	52.6	50.1	57.6	51.8	$\underset{=}{39.5}$
CSK	35.0	36.9	38.6	36.5	42.1	34.3	30.5	31.6	39.9	34.9	35.0
KCF	42.7	49.3	49.5	51.4	53.5	53.4	49.7	45.9	49.7	55.0	31.2
CN	38.4	41.4	44.1	42.5	44.9	43.4	41.0	37.3	46.9	41.0	31.1

Table 3. Performance of seven algorithms on 11 challenging attributes in OTB-2013 datasetunit: %

Algorithm	Running speed /(frame·s^-1)
Algorithm	OTB-2013	OTB-100
Ours	37.8657	37.3656
ECO-HC	35.5110	34.6002
BACF	27.0998	27.4870
SRDCFdecon	2.4089	2.2993
CCOT	0.2279	0.2160
Staple	44.9039	42.8838
SRDCF	3.6016	3.5327
CF2	11.0193	10.4229
KCF	127.1455	113.4207
DSST	32.3246	25.2587
SAMF	18.5867	16.9518

Table 4. Comparison of running speed of 11 algorithms on OTB-2013 and OTB-100 test datasets

Haifeng Liu, Cheng Sun, Xingliang Liang. Correlation-Filter Tracking Algorithm with Adaptive-Feature Fusion and Anti-Occlusion[J]. Laser & Optoelectronics Progress, 2020, 57(14): 141014

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information