Multi‐Object Tracking Based on YOLOv8 and Quasi‐Dense Similarity Learning

Wenwu Cao; Daqun Li; Zhijia Wu; Jianzhuo Liu

doi:10.3788/CJL241243

Chinese Journal of Lasers, Volume. 52, Issue 6, 0604003(2025)

Multi‐Object Tracking Based on YOLOv8 and Quasi‐Dense Similarity Learning

Wenwu Cao, Daqun Li^*, Zhijia Wu, and Jianzhuo Liu

Precision Instrument and Equipment R & D Center, Changchun Institute of Optics, Fine Mechanics and Physics, Chinese Academy of Sciences, Changchun 130033, Jilin , China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(12)

Fig. 1. Overall framework of multi- object tracking

Download full size

Fig. 2. Network Framework of YOLOv8

Download full size

Fig. 3. CBAM framework

Download full size

Fig. 4. Improved YOLOv8 network framework with CBAM introduced

Download full size

Fig. 5. Similarity learning framework of adjacent keyframe. (a) Key image; (b) reference image

Download full size

Fig. 6. AFLink framework

Download full size

Fig. 7. Tracking effects of module ablation experiment. (a)‒(c) Before model improvement; (d)‒(f) after model improvement

Download full size

Fig. 8. Experimental test results of multiple airplane images. (a) 71^st frame image in video clip 1; (b) 236^th frame image in video clip 1; (c) 56^th frame image in video clip 2; (d) 189^th frame image in video clip 2

Download full size

Table 1. MOT evaluation indicators

View table

Table 1. MOT evaluation indicators

Indicator	Definition
MOTA （↑）	Multiple object tracking accuracy
IDF1 （↑）	Identity F1 score
HOTA （↑）	High order tracking accuracy
ID1 （↓）	Number of changes in target tracking ID
FN （↓）	Number of missed detections
FP （↓）	Number of false detections
FPS （↑）	frame rate

Table 2. Results of module ablation experiments
View table
Table 2. Results of module ablation experiments
QDSL SoftMax CBAM GSI AFLink MOTA/% IDF1/%
76.9 85.7
√ 77.1 85.8
√ √ 77.5 86.0
√ √ √ 80.2 86.7
√ √ √ √ 80.5 87.2
√ √ √ √ √ 80.7 87.4

Table 3. Comparative experimental results under different threshold parameters

View table

Table 3. Comparative experimental results under different threshold parameters

$β_{o b j}$	$β_{m a t c h}$	$β_{n e w}$	MOTA /%	IDF1 /%
0.85	0.90	0.98	80.65	87.37
0.80	0.90	0.98	80.58	87.38
0.90	0.90	0.98	80.57	87.31
0.85	0.85	0.98	80.62	87.32
0.85	0.85	0.98	80.65	87.34
0.85	0.90	0.95	80.70	87.40
0.85	0.90	1.00	80.67	87.38
0.80	0.85	0.95	80.70	87.37
0.90	0.90	1.00	80.68	87.40

Table 4. Comparative experimental results of different algorithms
View table
Table 4. Comparative experimental results of different algorithms
Algorithm MOTA /% IDF1 /% ID1 FP FN
ByteTrack 70.2 79.3 21 406 4175
SORT 70.5 80.6 19 432 4296
deepSORT 71.3 82.4 14 986 3592
Ours 72.4 83.0 11 749 2841

Tools

Get Citation

Copy Citation Text

Wenwu Cao, Daqun Li, Zhijia Wu, Jianzhuo Liu. Multi‐Object Tracking Based on YOLOv8 and Quasi‐Dense Similarity Learning[J]. Chinese Journal of Lasers, 2025, 52(6): 0604003

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites