Lightweight Network for Real-Time Object Detection in Fisheye Cameras

Xinlei Wang; Chenxu Liao; Shuo Wang; Ruilin Xiao

doi:10.3788/LOP241236

Journals >Laser & Optoelectronics Progress >Volume 62 >Issue 2 >Page 0237012 > Article

Laser & Optoelectronics Progress
Vol. 62, Issue 2, 0237012 (2025)

Lightweight Network for Real-Time Object Detection in Fisheye Cameras

Xinlei Wang^1、2、*, Chenxu Liao¹, Shuo Wang¹, and Ruilin Xiao¹

Author Affiliations

¹School of Electronic and Information Engineering, Nanjing University of Information Science and Technology, Nanjing 210044, Jiangsu , China

²School of Electronic Information Engineering, Wuxi University, Wuxi 214105, Jiangsu , China

show less

DOI: 10.3788/LOP241236 Cite this Article Set citation alerts

Xinlei Wang, Chenxu Liao, Shuo Wang, Ruilin Xiao. Lightweight Network for Real-Time Object Detection in Fisheye Cameras[J]. Laser & Optoelectronics Progress, 2025, 62(2): 0237012 Copy Citation Text

show less

Fig. 1. Overall framework of Fisheye-YOLOv8

Download full size

Fig. 2. Faster-EMA module and convolution calculation process

Download full size

Fig. 3. Calculation process of EMA module

Download full size

Fig. 4. The fusion mode of BiFPN

Download full size

Fig. 5. Internal structure of RFAConv

Download full size

Fig. 6. RetinaNet architecture

Download full size

Fig. 7. Comparison of G-LHead and original detection head modifications

Download full size

Fig. 8. The detection effects under different conditions. (a)‒(c) Original images; (a1)‒(c1) detection results of EfficientDet; (a2)‒(c2) detection results of PGDS-YOLOv8s; (a3)‒(c3) detection results of YOLOv8m; (a4)‒(c4) detection results of Fisheye-YOLOv8

Download full size

Fig. 9. Front camera and side camera detection results.(a)‒(b) Original images; (a1)‒(b1) detection results of EfficientDet; (a2)‒(b2) detection results of PGDS-YOLOv8s; (a3)‒(b3) detection results of YOLOv8m; (a4)‒(b4) detection results of Fisheye-YOLOv8

Download full size

Fig. 10. Detection effects of LOAF dataset. (a)‒(d) Original images; (a1)‒(d1) detection results of YOLOv8m; (a2)‒(d2) detection results of Fisheye-YOLOv8

Download full size

Fig. 11. Detection effects of VisDrone2019 dataset. (a)‒(d) Original image; (a1)‒(d1) detection results of YOLOv8m; (a2)‒(d2) detection results of Fisheye-YOLOv8

Download full size

Method	P /%	R /%	mAP /%	Params /10⁶	GFLOPs /10⁹	FPS
YOLOv8m	79.8	34.6	58.3	25.8	78.7	89
YOLOv8m+Faster-EMA	79.5	34.4	58.3	20.4	68.2	95
YOLOv8m+Faster-EMA+RFA-BiFPN	79.9	35.8	59.9	12.2	61.2	101
YOLOv8m+Faster-EMA+RFA-BiFPN+G-LHead	82.2	34.7	59.8	10.0	48.7	109
YOLOv8m+Faster-EMA+RFA-BiFPN+G-LHead+LAMP	83.1	36.3	60.5	5.3	23.4	118

Table 1. The impact of each module on different indicators

Model	mAP /%
Model	bus	bike	car	pedestrian	truck
YOLOv8m	69.5	58.3	69.7	36.5	44.9
YOLOv8m+Faster-EMA	68.1	57.9	70.5	37.1	45.5
YOLOv8m+Faster-EMA+RFA-BiFPN	67.8	58.8	71.1	41.8	59.6
YOLOv8m+Faster-EMA+RFA-BiFPN+G-LHead	66.3	59.1	71.2	43.1	59.2
YOLOv8m+Faster-EMA+RFA-BiFPN+G-LHead+LAMP	70.1	59.8	70.6	43.2	58.8

Table 2. Influences of each module on different detection objects

Model	P /%	R /%	mAP /%	Params /10⁶	GFLOPs /10⁹	FPS
Faster-RCNN	65.2	27.1	24.8	136.9	370.2	27
EfficientDet	69.4	33.3	44.5	12.0	25.1	29
SSD	67.8	32.1	37.3	26.3	62.7	43
YOLOv5s	81.5	25.9	54.1	9.1	23.8	123
YOLOv7	78.9	41.5	39.9	36.5	103.2	129
YOLOv8m	79.8	34.6	58.3	25.8	78.7	89
FisheyeDet	78.2	33.5	55.3	16.3	44.2	108
PGDS-YOLOv8s	82.3	35.7	60.2	10.8	28.5	115
Fisheye-YOLOv8	83.1	36.3	60.5	5.3	23.4	118

Table 3. Performance comparison of different algorithms on Fisheye8K data set

Model	P /%	R /%	mAP /%
Faster-RCNN	63.7	25.3	23.1
EfficientDet	67.9	38.5	42.5
SSD	68.6	40.6	38.6
YOLOv5s	78.3	35.2	56.6
YOLOv7	75.7	46.2	42.4
YOLOv8m	76.5	41.2	58.5
FisheyeDet	76.8	43.1	56.8
PGDS-YOLOv8s	78.2	42.3	59.2
Fisheye-YOLOv8	78.5	44.2	59.7

Table 4. Performance comparison of different algorithms on WoodScape datasets

Model	P /%	R /%	mAP /%	Params /10⁶	GFLOPs /10⁹	FPS
YOLOv8m	77.5	40.5	61.2	25.8	78.7	95
Fisheye-YOLOv8	77.3	45.3	62.3	5.3	23.4	127

Table 5. Generalization studies on LOAF datasets

Model	P /%	R /%	mAP /%	Params /10⁶	GFLOPs /10⁹	FPS
YOLOv8m	51.2	40.5	41.1	25.8	78.7	98
Fisheye-YOLOv8	53.4	41.0	42.6	5.3	23.4	132

Table 6. Generalization study on VisDrone2019 dataset

Xinlei Wang, Chenxu Liao, Shuo Wang, Ruilin Xiao. Lightweight Network for Real-Time Object Detection in Fisheye Cameras[J]. Laser & Optoelectronics Progress, 2025, 62(2): 0237012

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information