Tracking Algorithm for Siamese Network Based on Target-Aware Feature Selection

Zhiwang Chen; Zhongxin Zhang; Juan Song; Hongfu Luo; Yong Peng

doi:10.3788/AOS202040.0915003

Acta Optica Sinica, Volume. 40, Issue 9, 0915003(2020)

Tracking Algorithm for Siamese Network Based on Target-Aware Feature Selection

Zhiwang Chen^1,2, Zhongxin Zhang^1、*, Juan Song³, Hongfu Luo¹, and Yong Peng⁴

¹Key Laboratory of Industrial Computer Control Engineering of Hebei Province, Yanshan University, Qinhuangdao, Hebei 0 66004, China

²National Engineering Research Center for Equipment and Technology of Cold Strip Rolling, Yanshan University, Qinhuangdao, Hebei 0 66004, China

³Jiamusi Electric Power Company, State Grid Heilongjiang Electric Power Co., Ltd., Jiamusi, Heilongjiang 154002, China

⁴School of Electrical Engineering, Yanshan University, Qinhuangdao, Hebei 0 66004, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(15)

Fig. 1. Development process diagram of object tracking algorithm based on Siamese network

Download full size

Fig. 2. Framework of tracking algorithm with Siamese network based on target-aware feature selection

Download full size

Fig. 3. Comparison experiment results of original features of each layer in the feature extraction module

Download full size

Fig. 4. Visualized results of the original features and the target-aware features. (a)(b) Template frame and detection frame; (c)(f)(g) feature maps of layer3; (d)(h)(i) feature maps of layer4; (e)(j)(k) feature maps of layer5

Download full size

Fig. 5. Visualized results of the original features and target-aware features. (a) Original image; (b) layer3; (c) layer4; (d) layer5

Download full size

Fig. 6. Specific structure diagram of SiamRPN module

Download full size

Fig. 7. Visualization of anchor bounding boxes. (a) Visualization of 25×25×K anchor bounding boxes; (b) visualization of 5 anchor bounding boxes at center position of target

Download full size

Fig. 8. Flow chart of proposed algorithm

Download full size

Fig. 9. Comparison of success rate and precision plots of OPE for realtime trackers on the OTB100 dataset. (a) Success rate plot;(b) precision plot

Download full size

Fig. 10. Success rate plots with 11 different attributes for 8 trackers. (a) Low resolution; (b) out of plane rotation; (c) fast motion; (d) out of view; (e) scale variation; (f) motion blur; (g) in-plane rotation; (h) deformation; (i) illumination variation; (j) occlusion; (k) background clutters

Download full size

Fig. 11. Actual tracking results of each algorithm for videos with different attributes

Download full size

Table 1. Comparison of experimental results of two template frame feature maps on the OTB2015 dataset
View table
Table 1. Comparison of experimental results of two template frame feature maps on the OTB2015 dataset
Frame Success Precision
Template frame 1 0.661 0.878
Template frame 2 0.547 0.786

Table 2. Comparison of experimental results for changing the number of important feature channels n₄ in layer 4 on the OTB2015 dataset (n₃=256, n₅=1500)

View table

Table 2. Comparison of experimental results for changing the number of important feature channels n₄ in layer 4 on the OTB2015 dataset (n₃=256, n₅=1500)

Video sequence	Success, precision				SiamRPN++
Video sequence	n₄=512	n₄=650	n₄=800	n₄=1024	SiamRPN++
Basketball	0.420,0.532	0.432,0.548	0.436,0.557	0.425,0.584	0.446,0.562
Bird2	0.708,0.827	0.701,0.825	0.688,0.813	0.698,0.825	0.627,0.748
Bird1	0.176,0.497	0.211,0.560	0.203,0.448	0.236,0.502	0.204,0.367
Bolt	0.261,0.335	0.259,0.337	0.257,0.340	0.650,0.883	0.644,0.887
Girl2	0.614,0.720	0.586,0.685	0.557,0.645	0.579,0.679	0.634,0.720
Car4	0.838,0.953	0.850,0.954	0.849,0.954	0.847,0.951	0.869,0.953
ClifBar	0.316,0.396	0.605,0.836	0.577,0.819	0.567,0.795	0.524,0.718
Dancer	0.779,0.871	0.763,0.852	0.735,0.829	0.741,0.831	0.768,0.861
DragonBaby	0.626,0.748	0.629,0.750	0.630,0.743	0.630,0.747	0.681,0.830
FaceOcc1	0.637,0.534	0.636,0.528	0.608,0.560	0.621,0.559	0.604,0.486
Freeman3	0.811,0.954	0.814,0.955	0.816,0.955	0.815,0.954	0.809,0.960
Human2	0.743,0.773	0.756,0.782	0.768,0.799	0.782,0.806	0.775,0.817
Jumping	0.550,0.804	0.600,0.840	0.610,0.845	0.578,0.816	0.670,0.882
Liquor	0.750,0.814	0.708,0.770	0.701,0.764	0.607,0.653	0.616,0.661
Suv	0.745,0.901	0.736,0.898	0.679,0.882	0.434,0.519	0.649,0.802
Woman	0.668,0.901	0.673,0.903	0.670,0.899	0.650,0.890	0.613,0.906

Table 3. Comparison of experimental results on the OTB2015 dataset

View table

Table 3. Comparison of experimental results on the OTB2015 dataset

Tracker name	Success	Precision	V_FPS
SiamRPN++	0.695	0.905	35
Ta-SiamRPN++	0.661	0.878	36
SiamRPN	0.643	0.860	71
RASNet^[11]	0.642		83
SA-Siam^[25]	0.657	0.865	50
CFNet^[26]	0.568	0.748	75
SiamFC	0.582	0.771	49
TADT	0.647	0.839	34
DaSiamRPN	0.658	0.880	97
BACF^[27]	0.617	0.815	35
ECO	0.694	0.910	3
UPDT	0.702		0.4
STRCF^[28]	0.683		3

Table 4. Comparison of experimental results on the VOT2018 dataset

View table

Table 4. Comparison of experimental results on the VOT2018 dataset

Tracker name	Accuracy	Robustness	EAO	Lostnumber	V_FPS
SiamRPN++	0.601	0.234	0.415	50	35
DaSiamRPN	0.586	0.276	0.383	59	59
UPDT	0.536	0.184	0.378	39	0.4
RCO^[30]	0.507	0.155	0.376	33	0.8
DeepSTRCF^[31]	0.523	0.215	0.345	46	3
SA_Siam_R	0.566	0.258	0.337	55	32
SiamVGG^[32]	0. 531	0.318	0.286	68	29
ECO	0.484	0.276	0.280	59	4
DSiam^[33]	0.512	0.654	0.196	138	10
SiamFC	0.503	0.585	0.187	125	32
DCFNet^[34]	0.470	0.543	0.182	116	27
DensSiam^[35]	0.462	0.688	0.174	147	19
Ta-SiamRPN++	0.593	0.272	0.360	58	36

Tools

Get Citation

Copy Citation Text

Zhiwang Chen, Zhongxin Zhang, Juan Song, Hongfu Luo, Yong Peng. Tracking Algorithm for Siamese Network Based on Target-Aware Feature Selection[J]. Acta Optica Sinica, 2020, 40(9): 0915003

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Machine Vision

Received: Dec. 19, 2019

Accepted: Jan. 19, 2020

Published Online: May. 6, 2020

The Author Email: Zhongxin Zhang (ZZXin00016@163.com)

DOI:10.3788/AOS202040.0915003

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology

Table 1. Comparison of experimental results of two template frame feature maps on the OTB2015 dataset

Table 1. Comparison of experimental results of two template frame feature maps on the OTB2015 dataset

Table 2. Comparison of experimental results for changing the number of important feature channels n4 in layer 4 on the OTB2015 dataset (n3=256, n5=1500)

Table 2. Comparison of experimental results for changing the number of important feature channels n4 in layer 4 on the OTB2015 dataset (n3=256, n5=1500)

Table 3. Comparison of experimental results on the OTB2015 dataset

Table 3. Comparison of experimental results on the OTB2015 dataset

Table 4. Comparison of experimental results on the VOT2018 dataset

Table 4. Comparison of experimental results on the VOT2018 dataset

Table 2. Comparison of experimental results for changing the number of important feature channels n₄ in layer 4 on the OTB2015 dataset (n₃=256, n₅=1500)

Table 2. Comparison of experimental results for changing the number of important feature channels n₄ in layer 4 on the OTB2015 dataset (n₃=256, n₅=1500)