Target Tracking Algorithm Based on Siamese Network of Feature Optimization Model

Yongqiang Wu; Baohua Zhang; Xiaoqi Lv; Yu Gu; Yueming Wang; Xin Liu; Yan Ren; Jianjun Li; Ming Zhang

doi:10.3788/LOP202259.1215003

Journals >Laser & Optoelectronics Progress >Volume 59 >Issue 12 >Page 1215003 > Article

Laser & Optoelectronics Progress
Vol. 59, Issue 12, 1215003 (2022)

Target Tracking Algorithm Based on Siamese Network of Feature Optimization Model

Yongqiang Wu¹, Baohua Zhang^1、3、*, Xiaoqi Lv^2、3, Yu Gu^1、3, Yueming Wang^1、3, Xin Liu^1、3, Yan Ren¹, Jianjun Li^1、3, and Ming Zhang^1、3

Author Affiliations

¹School of Information Engineering, Inner Mongolia University of Science and Technology, Baotou 014010, Inner Mongolia , China

²School of Information Engineering, Mongolia Industrial University, Huhehaote010051, Inner Mongolia , China

³Inner Mongolia Key Laboratory of Patten Recognition and Intelligent Image Processing, Baotou 014010, Inner Mongolia , China

show less

DOI: 10.3788/LOP202259.1215003 Cite this Article Set citation alerts

Yongqiang Wu, Baohua Zhang, Xiaoqi Lv, Yu Gu, Yueming Wang, Xin Liu, Yan Ren, Jianjun Li, Ming Zhang. Target Tracking Algorithm Based on Siamese Network of Feature Optimization Model[J]. Laser & Optoelectronics Progress, 2022, 59(12): 1215003 Copy Citation Text

show less

Fig. 1. Model frame diagram

Download full size

Fig. 2. Hourglass network and feature optimization model

Download full size

Fig. 3. Channel attention module^[5]

Download full size

Fig. 4. Spatial attention module^[5]

Download full size

Fig. 5. Training process diagram of loss function

Download full size

Fig. 6. Comparison of test results of various algorithms in OTB100 data set. (a) Precision rate; (b) success rate

Download full size

Fig. 7. Test results of various algorithms on Soccer sequence

Download full size

Fig. 8. Test results of various algorithms on MotorRolling sequence

Download full size

Fig. 9. Test results of various algorithms on Jogging sequence

Download full size

Fig. 10. Results of ablation experiment. (a) Precision rate; (b) Success rate

Download full size

Layer number	Network structure	Convolution kernels	Stride	Channel number	Template image /pixel	Search image /pixel
Layer1	Input layer			‒3	135×135	263×263
	Conv2d	3	1	96‒3	133×133	261×261
	Conv2d	3	1	96‒96	131×131	259×259
	MaxPool2d	3	2	‒	65×65	129×129
Layer2	Conv2d	3	1	128‒96	63×63	127×127
	Conv2d	3	1	128‒128	61×61	125×125
	MaxPool2d	3	2	‒	30×30	62×62
Layer3	Conv2d	3	1	256‒128	28×28	60×60
	Conv2d	3	1	256‒256	26×26	58×58
	Conv2d	3	1	256‒256	24×24	56×56
	MaxPool2d	2	2	‒	12×12	28×28
Layer4	Conv2d	3	1	512‒256	10×10	26×26
Layer5	Conv2d	3	1	512‒512	8×8	24×24

Table 1. Structure of backbone network

Name	Epoch	Got-10k	ILSVR2015_VID	Precision rate	Success rate	Speed/（frame·s^-1）
SCSAtt	50	√	√	0.855	0.641	59.871
Proposed	20	√	√	0.853	0.648	59.497

Table 2. Comparison of experimental data

Name	Accuracy	EAO	Speed /（frame·s^-1）
Proposed	0.5360	0.1920	44.33
SiamFC	0.4943	0.1875	31.89
LSART	0.4932	0.3230	1.00
CSRDCF	0.4910	0.2562	10.20
DeepSRDCF	0.4896	0.2753	65.30
ECO-HC	0.4842	0.2486	75.60

Table 3. Comparison of test results of various algorithms in VOT2018 data set

Name		IPR	IV	BC	OCC	DEF	SV	LR	FM	OPR	OV	MB
Proposed	Suc	0.624	0.646	0.609	0.613	0.609	0.636	0.682	0.616	0.630	0.545	0.628
Proposed	Pre	0.842	0.844	0.808	0.807	0.831	0.846	0.998	0.797	0.854	0.715	0.800
Siam RPN	Suc	0.628	0.649	0.591	0.585	0.617	0.615	0.639	0.599	0.625	0.542	0.622
Siam RPN	Pre	0.854	0.859	0.799	0.780	0.825	0.838	0.978	0.789	0.851	0.726	0.816
Siam DWfc	Suc	0.606	0.622	0.574	0.601	0.560	0.613	0.596	0.630	0.612	0.590	0.654
Siam DWfc	Pre	0.824	0.794	0.762	0.798	0.763	0.819	0.901	0.808	0.829	0.781	0.841
CFNet	Suc	0.567	0.541	0.561	0.527	0.526	0.546	0.614	0.554	0.553	0.454	0.540
CFNet	Pre	0.786	0.707	0.756	0.699	0.714	0.731	0.888	0.705	0.759	0.601	0.680
Siam FC	Suc	0.559	0.572	0.527	0.549	0.512	0.556	0.618	0.571	0.561	0.509	0.554
Siam FC	Pre	0.743	0.736	0.692	0.723	0.691	0.736	0.900	0.744	0.758	0.673	0.707
Staple	Suc	0.548	0.529	0.560	0.543	0.551	0.521	0.394	0.540	0.533	0.475	0.541
Staple	Pre	0.768	0.783	0.749	0.726	0.752	0.726	0.690	0.708	0.737	0.664	0.698
SRDCF	Suc	0.544	0.613	0.583	0.559	0.544	0.561	0.514	0.597	0.550	0.460	0.594
SRDCF	Pre	0.745	0.792	0.775	0.734	0.734	0.745	0.760	0.768	0.741	0.594	0.765
fDSST	Suc	0.505	0.559	0.523	0.460	0.427	0.475	0.382	0.458	0.477	0.386	0.469
fDSST	Pre	0.698	0.722	0.704	0.602	0.550	0.648	0.678	0.570	0.654	0.474	0.566

Table 4. Challenge performance results of various algorithms in OTB100 data set

Name	Improved VGG-Net	Improved Hourglass	AlexNet	Precision rate	Success rate
Proposed-2	√	√	-	0.560	0.724
Proposed-A	-	√	√	0.538	0.703

Table 5. Overall data of deep network ablation experiment

Name		IPR	IV	BC	OCC	DEF	SV	LR	FM	OPR	OV	MB
Proposed-2	Suc	0.522	0.510	0.479	0.501	0.490	0.543	0.604	0.559	0.529	0.439	0.544
Proposed-2	Pre	0.677	0.642	0.619	0.643	0.647	0.709	0.872	0.695	0.700	0.569	0.664
Proposed-A	Suc	0.511	0.480	0.458	0.487	0.462	0.520	0.554	0.536	0.522	0.431	0.532
Proposed-A	Pre	0.659	0.608	0.614	0.624	0.622	0.686	0.812	0.673	0.691	0.571	0.663

Table 6. Experimental data of OTB100 challenge ablation in deep network

Name	Improved VGG-Net	Improved Hourglass	Hourglass	Layer	Precision rate	Success rate
SCSAtt	√	-	-	-	0.687	0.529
Proposed-1	√	√	-	1	0.698	0.538
Proposed-2	√	√	-	2	0.724	0.560
Proposed-3	√	√	-	3	0.704	0.539
Proposed-No	√	-	√	2	0.694	0.530

Table 7. Ablation experiment data

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information