A Scale Adapted Tracking Algorithm Based on Kernelized Correlation

Xiufeng Liao; Zhiqiang Hou; Wangsheng Yu; Jiaoyao Wang; Chuanhua Chen

doi:10.3788/AOS201838.0715002

Journals >Acta Optica Sinica >Volume 38 >Issue 7 >Page 0715002 > Article

Acta Optica Sinica
Vol. 38, Issue 7, 0715002 (2018)

A Scale Adapted Tracking Algorithm Based on Kernelized Correlation

Xiufeng Liao¹, Zhiqiang Hou^1、2、*, Wangsheng Yu¹, Jiaoyao Wang¹, and Chuanhua Chen¹

Author Affiliations

¹ Information and Navigation Institute of Air Force Engineering University, Xi'an, Shaanxi 710077, China

² School of Computer Science & Technology, Xi'an University of Posts & Telecommunications, Xi'an,Shaanxi 710121, China;

show less

DOI: 10.3788/AOS201838.0715002 Cite this Article Set citation alerts

Xiufeng Liao, Zhiqiang Hou, Wangsheng Yu, Jiaoyao Wang, Chuanhua Chen. A Scale Adapted Tracking Algorithm Based on Kernelized Correlation[J]. Acta Optica Sinica, 2018, 38(7): 0715002 Copy Citation Text

show less

Fig. 1. Extraction process of one-dimensional features

Download full size

Fig. 2. Flow chart of proposed tracking method

Download full size

Fig. 3. Qualitative comparison of tracking results of 9 trackers

Download full size

Fig. 4. (a) Precision plots and (b) success plots of 28 sequences with scale variations

Download full size

Fig. 5. (a) Precision plots and (b) success plots of 51 sequences

Download full size

Input: Image sequence: I₁, I₂, …, I_n. Initial target position: p₀=(x₀,y₀), and initial target scale: s₀=(w₀,h₀)
Output: The estimated position of target: p_t=(x_t,y_t), and estimated scale: s_t=(w_t, h_t)
for t=1,2,3,…, n, do:
1	Locate the ROI area in frame # t centered at p_t_-1 with the scale of s_t_-1;
2	Crop out the ROI image and resize to the size of sample template;
3	Extract the HOG and color features;
4	Learn the kernelized correlation response map using Eq.(10);
5	Locate the center of the target p_t in frame # t using Eq.(12);
6	Obtain the multi-scale sample image I_s={I_s₁,I_s₂,…,I_sS} in frame # t based on p_t and s_t-₁;
7	Build scale filters by extracting fusion features from the above multi-scale image;
8	Compute the kernelized correlation response score using Eq.(16);
9	Estimate the optimal scale s_t of target in the frame # t using Eq.(17) and Eq.(18);
10	Update the translation filters using Eq.(19) and Eq.(20);
11	Update the scale filters using Eq.(21).
Until	End of the image sequence.

Table 1. Scale adapted tracking algorithm based on kernelized correlation

Sequence	Proposed	CNT	DSST	CST	SST	KCF	NRMLC	DLT	CN
car scale	7.3(75)	23.4(57)	18.8(46)	12.5(48)	87(59)	16.1(47)	10.4(74)	25.5(61)	25.2(43)
fleet face	22(69)	24.7(60)	28.3(58)	67.2(54)	60.8(67)	26.4(63)	96.9(44)	27.5(56)	126.2(26)
dog1	3.8(99)	6.8(95)	4.6(66)	4.8(67)	4.9(89)	4.1(64)	11.6(91)	4.4(92)	3.5(68)
singer2	7.2(100)	6.8(100)	8.2(100)	7.4(100)	175.2(4)	10.2(100)	195.4(3)	173.0(3)	167.1(4)
skating1	6.2(53)	6.7(60)	6.8(52)	7.9(58)	8.8(62)	7.7(50)	15(52)	52.9(49)	8.0(52)
shaking	7.7(74)	74.1(5)	8(73)	5.7(72)	8.1(75)	113.2(4)	109.4(3)	-	15.1(60)
sylvester	8.7(75)	10.7(62)	14.8(63)	12.9(68)	11.2(63)	13.3(67)	24.5(49)	10.9(51)	9.5(68)
basketball	5.3(78)	534.1(6)	111.6(28)	23.5(57)	106.0(22)	8.1(67)	60.0(7)	12.0(51)	9.3(63)
tiger1	12(68)	94.2(15)	19.5(63)	11.2(74)	93.5(16)	15.7(68)	54.5(22)	23.2(58)	61.2(20)
freeman1	7.4(59)	7.9(53)	112.5(24)	9.7(41)	9.8(37)	94.6(23)	7.4(49)	103.6(28)	159.9(22)
coke	10.5(65)	36.7(35)	12.7(60)	148.7(4)	25.9(45)	18.7(55)	62.1(17)	20.1(53)	30.8(42)
jogging-1	4.3(80)	6.2(62)	112(19)	3.9(81)	144.6(20)	87.9(19)	7.2(75)	113(18)	101.7(19)
Average	8.5(75)	69.4(51)	38.2(55)	26.3(61)	61.3(47)	34.7(52)	54.5(40)	51.5(45)	59.8(41)

Table 2. Center location errors and overlap rates of 12 sequences

Algorithm	SV(28)	IV(25)	OCC(29)	BC(21)	DEF(19)	MB(12)	FM(17)	IPR(31)	OPR(39)	OV(6)	LR(4)
Proposed	0.744	0.757	0.792	0.756	0.831	0.624	0.630	0.779	0.791	0.648	0.516
CNT	0.662	0.566	0.662	0.646	0.687	0.507	0.500	0.661	0.672	0.502	0.557
DSST	0.740	0.741	0.725	0.691	0.657	0.603	0.562	0.780	0.732	0.533	0.534
CST	0.707	0.676	0.726	0.773	0.756	0.591	0.516	0.708	0.742	0.596	0.454
SST	0.688	0.603	0.588	0.644	0.487	0.408	0.425	0.630	0.599	0.406	0.527
KCF	0.680	0.729	0.749	0.752	0.741	0.650	0.602	0.725	0.730	0.649	0.379
NRMLC	0.597	0.437	0.583	0.497	0.541	0.378	0.397	0.511	0.546	0.492	0.542
DLT	0.590	0.534	0.574	0.495	0.563	0.453	0.446	0.548	0.561	0.444	0.396
CN	0.554	0.532	0.582	0.642	0.523	0.396	0.416	0.615	0.605	0.434	0.405

Table 3. Tracking precision values on 11 different attributes

Algorithm	SV(28)	IV(25)	OCC(29)	BC(21)	DEF(19)	MB(12)	FM(17)	IPR(31)	OPR(39)	OV(6)	LR(4)
Proposed	0.541	0.560	0.579	0.551	0.609	0.493	0.491	0.566	0.572	0.537	0.382
CNT	0.508	0.456	0.503	0.488	0.524	0.417	0.404	0.495	0.501	0.439	0.437
DSST	0.451	0.506	0.480	0.492	0.474	0.458	0.433	0.532	0.491	0.490	0.352
CST	0.466	0.486	0.506	0.567	0.551	0.474	0.411	0.496	0.514	0.509	0.349
SST	0.504	0.459	0.436	0.489	0.391	0.313	0.340	0.451	0.437	0.347	0.407
KCF	0.427	0.494	0.513	0.533	0.533	0.499	0.461	0.497	0.496	0.550	0.310
NRMLC	0.427	0.341	0.437	0.370	0.392	0.303	0.334	0.367	0.389	0.410	0.428
DLT	0.455	0.405	0.423	0.339	0.394	0.363	0.360	0.411	0.412	0.367	0.346
CN	0.363	0.390	0.404	0.453	0.388	0.329	0.334	0.437	0.418	0.410	0.311

Table 4. Tracking success rates on 11 attributes

Tracker	Proposed		CNT	DSST	CST	SST	KCF	NRMLC	DLT	CN
Code	M		M	M	M	M+C	M	M	M	M
Platform	CPU	CPU+GPU	CPU	CPU	CPU	CPU+GPU	CPU	CPU	CPU+GPU	CPU
Trackingspeed /(frame·s^-1)	43.2	56.3	5	24	2.2	2.2	172	1.28	15	-
Note: M, MATLAB; C, C++

Table 5. Tracking speed comparison of 9 trackers

Xiufeng Liao, Zhiqiang Hou, Wangsheng Yu, Jiaoyao Wang, Chuanhua Chen. A Scale Adapted Tracking Algorithm Based on Kernelized Correlation[J]. Acta Optica Sinica, 2018, 38(7): 0715002

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information