Foreground-Aware Based Spatiotemporal Correlation Filter Tracking Algorithm

Yueyang Yu; Zelin Shi; Yunpeng Liu

doi:10.3788/LOP56.221503

Journals >Laser & Optoelectronics Progress >Volume 56 >Issue 22 >Page 221503 > Article

Laser & Optoelectronics Progress
Vol. 56, Issue 22, 221503 (2019)

Foreground-Aware Based Spatiotemporal Correlation Filter Tracking Algorithm

Yueyang Yu^{1、2、3、4、5、*}, Zelin Shi^{1、2、3、4、5}, and Yunpeng Liu^{2、3、4、5}

Author Affiliations

¹School of Information Science and Technology, University of Science and Technology of China, Hefei, Anhui 230026, China

²Shenyang Institute of Automation, Chinese Academy of Sciences, Shenyang, Liaoning 110016, China

³Institutes for Robotics and Intelligent Manufacturing, Chinese Academy of Sciences, Shenyang, Liaoning 110016, China

⁴Key Laboratory of Opto-Electronic Information Processing, Chinese Academy of Science, Shenyang, Liaoning 110016, China

⁵Key Laboratory of Image Understanding and Computer Vision, Shenyang, Liaoning 110016, China

show less

DOI: 10.3788/LOP56.221503 Cite this Article Set citation alerts

Yueyang Yu, Zelin Shi, Yunpeng Liu. Foreground-Aware Based Spatiotemporal Correlation Filter Tracking Algorithm[J]. Laser & Optoelectronics Progress, 2019, 56(22): 221503 Copy Citation Text

show less

Fig. 1. Temporal consistency constraints with object area selection function explained by sequence Tiger

Download full size

Take one-dimensional vector as example, assuming length of target is D=3. Left side is one-dimensional signal xi with L=5. xi[Δτj] image is result of all cyclic shifts. Five one-dimensional vectors with length of 3 can be obtained by multiplying mask matrix P on this image, where first 3 rows are real positive samples with same size of object

Fig. 2. Take one-dimensional vector as example, assuming length of target is D=3. Left side is one-dimensional signal xi with L=5. xi[Δτj] image is result of all cyclic shifts. Five one-dimensional vectors with length of 3 can be obtained by multiplying mask matrix P on this image, where first 3 rows are real positive samples with same size of object

Download full size

Fig. 3. Comparison of training samples between traditional correlation filters and proposed method. (a) Cyclic-shift training samples of traditional correlation filter; (b) training samples of foreground-aware correlation filter

Download full size

Fig. 4. Relationship between IoU value and tracking confidence score for carRace and ball sequences without re-detector. (a) Relationship between IoU value of carRace and tracking confidence score; (b) 502nd-frame tracking result of carRace; (c) 510th-frame tracking result of carRace; (d) relationship between IoU of ball and tracking confidence score; (e) 209th-frame tracking result of ball; (f) 211st-frame tracking result of ball

Download full size

Fig. 5. Plots of OPE and success rate of trackers with traditional features on OTB-2013 dataset. (a) Plots of OPE; (b) plots of success rate

Download full size

Fig. 6. Plots of OPE and success rate of trackers with convolutional features on OTB-2013 dataset. (a) Plots of OPE; (b) plots of success rate

Download full size

Fig. 7. Comparison of tracking results of SiamFC, CCOT, DSST, KCF, ECO, CF2, and proposed algorithm on 8 challenging sequences from OTB-2015 dataset. From top to bottom: singer2, girl2, tiger, bird1, dragonbaby, motorrolling, skiing, and soccer

Download full size

Parameter	Ours	ECO-HC	LCT	SRDCF	Staple-CA	Staple	BACF	DSST	KCF
Mean OP /%	85.5	81.0	81.3	78.1	77.6	75.4	85.4	67.0	62.3
Mean DP /%	89.2	87.4	84.8	83.8	83.3	79.3	78.5	74.0	74.0
Tracking speed /(frame·s^-1)	25.3	42	18.5	5.8	35.3	76.6	23.2	20.4	171.8

Table 1. Success rate, precision, and tracking speed of tracking algorithm based on traditional features on OTB-2013 dataset

Algorithm	SV	OV	OR	OCC	DEF	MB	FM	IR	BC	LR	IV
ECO-HC	0.627	0.694	0.668	0.67	0.645	0.610	0.607	0.589	0.606	0.672	0.612
Ours	0.654	0.667	0.632	0.669	0.664	0.605	0.612	0.637	0.625	0.544	0.626
LCT	0.553	0.594	0.624	0.627	0.668	0.524	0.534	0.592	0.587	0.541	0.588
SRDCF	0.587	0.555	0.599	0.627	0.635	0.601	0.569	0.566	0.587	0.541	0.576
SAMF	0.507	0.555	0.559	0.612	0.625	0.461	0.483	0.525	0.520	0.526	0.513
Staple-CA	0.574	0.562	0.594	0.600	0.632	0.569	0.566	0.601	0.587	0.497	0.596
Staple	0.551	0.547	0.575	0.593	0.618	0.541	0.508	0.580	0.576	0.496	0.568
KCF	0.427	0.550	0.495	0.514	0.534	0.497	0.459	0.497	0.535	0.537	0.493
DSST	0.546	0.462	0.536	0.532	0.506	0.455	0.428	0.563	0.517	0.345	0.561

Table 2. Performance evaluation of each tracker on OTB-2013 dataset

Parameter	Ours	ECO	MDNet	CCOT	DeepSRDCF	SiamFC	CFNet	CF2
Mean OP /%	89.4	88.7	91.1	83.2	79.5	79.1	76.9	74.0
Mean DP /%	90.0	93.0	94.8	89.9	84.9	81.5	80.7	89.1
Tracking speed /(frame·s^-1)	10.6	9.8	0.8	0.8	0.2	83.7	78.4	10.2

Table 3. Success rate, precision, and tracking speed of tracking algorithm based on convolutional features on OTB-2013 dataset

Algorithm	EAO	Accuracy	Robustness
DSST	0.181	0.500	2.720
ECO	0.375	0.530	0.730
Staple	0.295	0.540	1.350
MDNet	0.257	0.530	1.200
BACF	0.223	0.560	1.880
SRDCF	0.247	0.520	1.500
ECO-HC	0.322	0.510	1.080
DeepSRDCF	0.276	0.510	1.170
CCOT	0.331	0.530	0.238
SiamFC	0.277	0.549	0.382
Ours	0.320	0.535	0.926
Oursdeep	0.285	0.555	1.330

Table 4. Evaluations of EAO, precision, and robustness of algorithms on VOT2016 dataset

Yueyang Yu, Zelin Shi, Yunpeng Liu. Foreground-Aware Based Spatiotemporal Correlation Filter Tracking Algorithm[J]. Laser & Optoelectronics Progress, 2019, 56(22): 221503

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information