Multitarget Tracking Algorithm Based on an Improved YOLOv3 Algorithm

Xiangsheng Zhang; Qing Shen

doi:10.3788/LOP202158.1610004

Laser & Optoelectronics Progress, Volume. 58, Issue 16, 1610004(2021)

Multitarget Tracking Algorithm Based on an Improved YOLOv3 Algorithm

Xiangsheng Zhang^* and Qing Shen

School of Internet of Things Engineering, Key Laboratory of Advanced Control of Light Industry Process, Ministry of Education, Jiangnan University, Wuxi, Jiangsu 214122, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(14)

Fig. 1. Structure of standard convolution filters

Download full size

Fig. 2. Structure of depthwise convolution filters

Download full size

Fig. 3. Structure of pointwise convolution filters

Download full size

Fig. 4. SENet structure

Download full size

Fig. 5. Improved YOLOv3 network model structure diagram

Download full size

Fig. 6. Multi-target tracking algorithm flow

Download full size

Fig. 7. Intersection-over-union of different number of anchors

Download full size

Fig. 8. Comparison of algorithm tracking results based on MOT15-PETS09 sequence. (a) YOLOv3-Deep-SORT tracking results; (b) our algorithm tracking results

Download full size

Fig. 9. Comparison of algorithm tracking results based on MOT16-06 sequence. (a) YOLOv3-Deep-SORT tracking results; (b) our algorithm tracking results

Download full size

Fig. 10. Comparison of algorithm tracking results based on ETHZ-eth02 sequence. (a) YOLOv3-Deep-SORT tracking results; (b) our algorithm tracking results

Download full size

Table 1. Size of a priori boxes with different numbers of a priori boxes

View table

Table 1. Size of a priori boxes with different numbers of a priori boxes

k=7	k=8	k=9	k=10	k=11
(18,69)	(17,70)	(16,69)	(18,40)	(18,72)
(26,82)	(23,62)	(18,40)	(18,75)	(20,40)
(28,64)	(25,81)	(20,80)	(23,64)	(23,80)
(33,96)	(31,72)	(23,64)	(25,83)	(26,64)
(34,74)	(33,92)	(26,82)	(31,73)	(28,80)
(41,85)	(38,77)	(31,72)	(33,99)	(34,89)
(49,113)	(43,94)	(33,94)	(37,86)	(34,72)
	(52,122)	(39,79)	(38,74)	(36,105)
		(46,107)	(46,101)	(38,85)
			(62,132)	(44,83)
				(51,114)

Table 2. Target detection algorithm performance comparison results
View table
Table 2. Target detection algorithm performance comparison results
Detection algorithm Avgmisrate /% F₁ /% FPS
Faster RCNN 32.15 88.57 5.52
YOLOv3 27.80 95.41 15.65
Our algorithm 13.60 96.56 22.35

Table 3. Comparison of the indicators of the test set on different sequences
View table
Table 3. Comparison of the indicators of the test set on different sequences
Sequence A_MOT /% ↑ P_MOT /% ↑ s_ID↓ FN↓
Venice-1 55.1 77.6 41 1212
KITTI-19 30.2 69.4 91 1626
KITTI-16 40.8 71.9 33 619
ETH-Crossing 66.3 80.4 15 252
PETS09-S2L2 55.6 73.1 177 2938
TUD-Crossing 76.8 72.8 21 202

Table 4. Comparison of evaluation indexes of multi-target tracking algorithms

View table

Table 4. Comparison of evaluation indexes of multi-target tracking algorithms

Algorithm	A_MOT /% ↑	P_MOT /% ↑	s_ID↓	FPS↑
YOLOv3-SORT	46.8	61.9	102	—
Faster RCNN-Deep-SORT	35.3	56.5	72	—
YOLOv3-Deep-SORT	54.8	68.0	68	2.8
YOLOv3-Kalman^[14]	39.2	66.2	107	—
SiamCNN^[15]	45.3	70.4	105	—
MOTDT^[22]	57.3	75.3	70	—
MDP^[23]	46.4	71.3	93	—
Our algorithm	56.0	78.2	57	4.4

Tools

Get Citation

Copy Citation Text

Xiangsheng Zhang, Qing Shen. Multitarget Tracking Algorithm Based on an Improved YOLOv3 Algorithm[J]. Laser & Optoelectronics Progress, 2021, 58(16): 1610004

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites