Attention-guided multi-scale infrared real-time detection of pedestrian and vehicle

Yinhui Zhang; Kai Ji; Zifen He; Guangchen Chen

doi:10.3788/IRLA20240063

Infrared and Laser Engineering, Volume. 53, Issue 5, 20240063(2024)

Attention-guided multi-scale infrared real-time detection of pedestrian and vehicle

Yinhui Zhang, Kai Ji, Zifen He^*, and Guangchen Chen

Author Affiliations

Mechanical and Electrical Engineering, Kunming University of Science and Technology, Kunming 650500, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(12)

Fig. 1. IRDet network model

Download full size

View in Article

Fig. 2. Visualization results of the height/width ratio of the truth bounding box

Download full size

View in Article

Fig. 3. Attention-guided global feature extraction module

Download full size

View in Article

Fig. 4. Cross-space perception module

Download full size

View in Article

$The distribution of \begin{document}$ \gamma $\end{document} before sparse training (a) and after sparse training (b)$

Fig. 5. The distribution of \begin{document}$ \gamma $\end{document} before sparse training (a) and after sparse training (b)

Download full size

View in Article

Fig. 6. Schematic diagram of channel pruning

Download full size

View in Article

Fig. 7. Comparision of detection effects. (a) Original image; (b) YOLOv5 s; (c) IRDet

Download full size

View in Article

Table 1. Anchor box size
View table
View in Article
Table 1. Anchor box size
Detect size Size Anchor size
128×128 Smaller [10, 15, 11, 30, 15, 50]
64×64 Small [19, 21, 25, 79, 32, 32]
32×32 Medium [42, 129, 50, 47, 74, 69]
16×16 Large [78, 181, 113, 99, 173, 171]

Table 2. Ablation experiment

View table

View in Article

Table 2. Ablation experiment

Model	K-Means++_4	AGFE	CSPM	AP-car	AP-person	mAP	Size/MB
YOLOv5s				86.9%	79.3%	83.1%	14.4
Proposed	√			87.9%	82.9%	85.4%	16.1
Proposed		√		88.4%	80.1%	84.3%	21.3
Proposed			√	88.1%	80.2%	84.2%	15.5
Proposed	√	√		89.4%	84.2%	86.8%	24.3
YOLO-IR	√	√	√	90.4%	85.6%	88.0%	25.5

Table 3. Scaling factor comparison experiment
View table
View in Article
Table 3. Scaling factor comparison experiment
Model mAP P/M Flops/G
YOLOv5s 83.1% 7.0 10.1
YOLOv5s_64 83.8% 10.4 15.5
YOLOv5s_32 84.3% 10.5 15.6
YOLOv5s_16 84.2% 10.5 15.6
YOLOv5s_8 84.3% 10.7 15.8

Table 4. Pruning experiment
View table
View in Article
Table 4. Pruning experiment
Pruning rate mAP Size/MB P/M
0 88.0% 25.5 12.3
0.7 87.5% 6.8 3.0
0.75 87.5% 6.3 2.7
0.8 87.4% 5.7 2.4
0.85 86.7% 5.4 2.2
0.9 84.5% 4.9 2.0

Table 5. Comparative experiment

View table

View in Article

Table 5. Comparative experiment

Model	AP-car	AP-person	mAP	Size/MB	P/M	Flops/G	FPS/frame·s⁻¹
SSD	80.6%	38.7%	59.6%	95.5	23.7	175.4	38
Faster-RCNN	65.1%	39.8%	52.5%	113.5	136.7	308.8	12
YOLOX-s	83.1%	78.0%	80.6%	71.8	8.9	17.1	138
YOLOv3	89.0%	80.6%	84.8%	123.5	61.5	98.9	60
YOLOv4	88.1%	82.7%	85.5%	105.5	52.4	76.2	66
YOLOv5-s	86.9%	79.3%	83.1%	14.4	7.0	10.1	150
YOLOv7-tiny	86.9%	79.0%	83.0%	12.3	6.0	8.4	131
Damoyolo	88.8%	80.4%	84.6%	65.5	15.6	35.6	93
YOLOv5-m	88.1%	80.5%	84.3%	42.1	20.8	30.6	98
YOLOv8-s	88.3%	83.4%	85.9%	22.5	11.1	18.2	122
IRDet	89.7%	85.0%	87.4%	5.7	2.4	8.7	110

Tools

Get Citation

Copy Citation Text

Yinhui Zhang, Kai Ji, Zifen He, Guangchen Chen. Attention-guided multi-scale infrared real-time detection of pedestrian and vehicle[J]. Infrared and Laser Engineering, 2024, 53(5): 20240063

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites