Multi-Task Learning Tracking Method Based on the Similarity of Dynamic Samples

Zaifeng Shi; Cheng Sun; Qingjie Cao; Zhe Wang; Qiangqiang Fan

doi:10.3788/LOP202158.1615002

Journals >Laser & Optoelectronics Progress >Volume 58 >Issue 16 >Page 1615002 > Article

Laser & Optoelectronics Progress
Vol. 58, Issue 16, 1615002 (2021)

Multi-Task Learning Tracking Method Based on the Similarity of Dynamic Samples

Zaifeng Shi^1、*, Cheng Sun^1、**, Qingjie Cao², Zhe Wang¹, and Qiangqiang Fan¹

Author Affiliations

¹School of Microelectronics, Tianjin University, Tianjin 300072, China

²School of Mathematical Sciences, Tianjin Normal University, Tianjin 300072, China

show less

DOI: 10.3788/LOP202158.1615002 Cite this Article Set citation alerts

Zaifeng Shi, Cheng Sun, Qingjie Cao, Zhe Wang, Qiangqiang Fan. Multi-Task Learning Tracking Method Based on the Similarity of Dynamic Samples[J]. Laser & Optoelectronics Progress, 2021, 58(16): 1615002 Copy Citation Text

show less

Fig. 1. Flow chart of the tracking method

Download full size

Fig. 2. Tracking results under different t_z values

Download full size

Fig. 3. Testing results of the OTB-2015 dataset. (a) Precision; (b) success rate

Download full size

Fig. 4. Tracking results of four video sequences. (a) Bird1; (b) Box; (c) Girl2; (d) Matrix

Download full size

Algorithm 1: Proposed tracking method
Input: Initial target position v₁, pretrained local tracker.
Output: Estimated target position v_q
1: Initial training the local tracker by the ﬁrst frame, set number of consecutive failures P_cf=0.
2: for q =2:m
3: if P_cf≤r then locally draw candidate regions of target
4: else globally draw candidate regions of target
5: Get the tracking result and confidence score using local tracker
6: if (c_on > 0 and s_cos>H_cos and q>h) or (c_on>0 and q≤h) then
P_cf =0, collect new short-term and long-term samples
7: else P_cf=P_cf +1
8: if (q % b≠0) then update local tracker using Eq. (4)
9: if (q % b==0) then update local tracker using Eq. (6)

Table 1. Implementation flow of the tracking method

Tracker	Detection of loss	Multi-task learning	Fast re-detection	Precision	Success rate
Baseline				0.9016	0.6677
Method 1	√			0.8986	0.6722
Method 2	√	√		0.9061	0.6788
Method 3	√		√	0.9032	0.6758
Method 4	√	√	√	0.9075	0.6812

Table 2. Results of ablation experiments

Tracker	IV	SV	OCC	DEF	MB	FM	IPR	OPR	OV	BC	LR
Proposed	0.909	0.889	0.858	0.876	0.854	0.877	0.907	0.898	0.838	0.937	0.937
MDNet	0.907	0.874	0.828	0.866	0.858	0.872	0.898	0.883	0.833	0.907	0.945
SiamRPN	0.859	0.838	0.780	0.826	0.816	0.789	0.854	0.851	0.725	0.799	0.978
SRDCF	0.792	0.745	0.734	0.735	0.765	0.768	0.745	0.741	0.593	0.775	0.760
SiamFC	0.736	0.736	0.723	0.691	0.707	0.744	0.743	0.758	0.672	0.692	0.900
Staple	0.783	0.726	0.728	0.752	0.698	0.708	0.768	0.737	0.664	0.749	0.690

Table 3. Tracking precision of 11 challengeable attributes

Tracker	C-COT	Staple	MDNet	SiamFC	SRDCF	Proposed
Accuracy	0.539	0.544	0.541	0.532	0.535	0.545
Robustness	0.238	0.378	0.337	0.461	0.419	0.317
AO	0.469	0.388	0.457	0.399	0.397	0.482

Table 4. Testing result of the VOT-2016 dataset

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information