Small-Target Traffic Sign Detection Based on Multiscale Feature Fusion

Experiment No.	Grid scale	Params /10⁷	P /%	R /%	mAP@0.5 /%	Speed /（frame·s^-1）
1	160×160， 80×80， 40×40， 20×20	7.54	88.0	80.3	85.2	56.8
2	160×160， 80×80， 40×40	5.71	88.0	79.8	85.8	62.5
3	80×80， 40×40， 20×20	5.75	87.6	80.1	84.9	45.7
4	80×80， 40×40	5.67	87.8	80.0	85.3	57.5
5	160×160， 80×80	5.65	85.6	82.4	85.9	63.3
6	160×160， 40×40	5.67	87.3	82.3	86.1	62.5

Table 2. Impact of different α and δ in Wise-IoU on the detection performance
View table
Table 2. Impact of different α and δ in Wise-IoU on the detection performance
（α， δ） Params /10⁷ P /% R /% mAP@0.5 /%
（2.5 ， 2） 5.67 88.2 79.6 86.3
（1.6 ， 4） 5.67 87.8 81.5 86.4
（1.4 ， 5） 5.67 88.5 81.7 86.5
（1.2 ， 6） 5.67 89.8 81.3 86.9
（1.0 ， 7） 5.67 89.2 79.9 86.5

Table 3. Ablation experiment

View table

Table 3. Ablation experiment

Method	Neck&Head	Wise-IoU v3	CoordConv	Params /10⁷	P /%	R /%	mAP@0.5 /%	Speed /（frame·s^-1）
YOLOv5s				7.1	86.8	79.2	85.5	64.1
A	√			5.7	87.3	82.3	86.1	62.5
B		√		7.1	85.9	76.3	85.8	64.1
C			√	10.3	87.1	80.2	85.7	56.2
D	√	√		5.7	89.8	81.3	86.9	62.5
E	√		√	6.2	89.1	81.7	87.2	55.7
F	√	√	√	6.2	88.8	82.3	87.7	55.7

Table 4. Comparison of different detection algorithms

View table

Table 4. Comparison of different detection algorithms

Method	Params	P /%	R /%	AP_small /%	mAP@0.5 /%	Speed /（frame·s^-1）
Faster R-CNN	144.5	58.8	76.5		72.1	3
Method in Ref.［16］	81.2	87.7	91.0		88.0
YOLOv3	61.5	72.6	78.0		83.4	49.3
YOLOv5s	7.1	86.8	79.2	46.9	85.5	64.1
YOLO7-tiny	6.2	87.3	81.7	43.7	82.9	79.5
YOLOv8n	3.0	78.5	70.5	46.4	77.8	90.9
YOLOv8s	11.2	85.6	76.8	50.2	85.8	52.3
STS-YOLO	6.2	88.8	82.3	51.5	87.7	55.7

Tools

Get Citation

Copy Citation Text

Fangke Jing, Hongge Ren, Song Li. Small-Target Traffic Sign Detection Based on Multiscale Feature Fusion[J]. Laser & Optoelectronics Progress, 2024, 61(12): 1237002

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites