Infrared Vehicle Detection Algorithm Based on Improved Shuffle-RetinaNet

Table 1. Experimental platform configuration
View table
Table 1. Experimental platform configuration
Platform and tools Version
Operating system Ubuntu 16.0.7
Graphic processing unit NVIDIA Quadro RTX 8000
Graphics driver 510.47.03
Programing language Python 3.7
Framework PyTorch 1.7.1

Table 2. Comparison of lightweight backbones

View table

Table 2. Comparison of lightweight backbones

Backbone	Number of parameters /10⁶	FLOPs /10⁹	Speed /（frame·s^-1）	$A P_{50}$ /%
ResNet50	36.10	65.35	14.2	87.4
ShuffleNetV1 1.0×（g=3）	11.91	27.32	29.8	84.9
ShuffleNetV2 1.0×（g=3）	11.15	24.09	32.4	86.9
MobileNetV2	14.01	26.97	28.4	85.1
MobileNetV3	14.98	24.03	30.3	86.1

Table 3. Overall ablation experimental results

View table

Table 3. Overall ablation experimental results

Method	DBAM	Improved feature network	Calibration factor	$A P_{50}$ /%	Number of parameters /10⁶	FLOPs /10⁹	Speed /（frame $∙ s^{- 1}$ ）
RetinaNet				87.4	36.10	65.35	14.2
Shuffle-RetinaNet				86.9	11.15	24.09	32.4
Ours	√			89.1	11.44	24.21	31.7
	√	√		91.1	11.74	24.35	31.1
	√	√	√	92.9	11.74	24.35	30.9

Table 4. Ablation study results of multi-scale detection accuracy
View table
Table 4. Ablation study results of multi-scale detection accuracy
Method DBAM Improved feature network Calibration factor $A P_{s}$ /% $A P_{m}$ /% $A P_{l}$ /%
RetinaNet 45.2 59.5 66.5
Shuffle-RetinaNet 42.9 55.7 64.7
Ours
√ 43.1 55.9 65.3
√ √ 44.9 58.4 65.9
√ √ √ 45.7 59.9 66.7

Table 5. Results of calibration factor comparison

View table

Table 5. Results of calibration factor comparison

Parameter value of calibration factor	AP /%	AP₅₀ /%
$α$ =0.5， $β$ =6	56.2	91.0
$α$ =0.5， $β$ =8	56.2	91.1
$α$ =0.5， $β$ =10	56.1	91.0
$α$ =1.0， $β$ =6	56.9	92.4
$α$ =1.0， $β$ =8	57.4	92.9
$α$ =1.0， $β$ =10	57.4	92.5
$α$ =1.5， $β$ =6	56.3	91.4
$α$ =1.5， $β$ =8	57.1	92.0
$α$ =1.5， $β$ =10	56.8	91.8

Table 6. Results of calibration factor comparison on Shuffle-RetinaNet

View table

Table 6. Results of calibration factor comparison on Shuffle-RetinaNet

Parameter value of calibration factor	Self-built dataset	FLIR ADAS
Parameter value of calibration factor	AP₅₀ /%	AP₅₀ /%
No calibration factor	86.9	87.6
$α$ =0.5， $β$ =6	87.1	87.9
$α$ =0.5， $β$ =8	87.2	87.8
$α$ =0.5， $β$ =10	87.1	87.9
$α$ =1.0， $β$ =6	88.9	89.5
$α$ =1.0， $β$ =8	89.1	89.7
$α$ =1.0， $β$ =10	89.0	89.5
$α$ =1.5， $β$ =6	88.6	89.1
$α$ =1.5， $β$ =8	88.7	89.2
$α$ =1.5， $β$ =10	88.4	89.0

Table 7. Comparision with classical object detection algorithms
View table
Table 7. Comparision with classical object detection algorithms
Method AP₅₀ /% Number of parameters /10⁶ FLOPs /10⁹ Speed /（frame $∙ s^{- 1}$ ）
Faster RCNN 88.6 41.12 67.75 12.4
RetinaNet 87.4 36.10 65.35 14.2
YOLOv5s 88.3 7.21 26.67 31.6
SSD512 86.2 36.04 60.85 18.1
Ours 92.9 11.74 24.35 30.9

Table 8. Comparision with classical infrared vehicle detection algorithms
View table
Table 8. Comparision with classical infrared vehicle detection algorithms
Method AP/% Number of parameters /10⁶
Algorithms 1^［20］ 76.57 20.60
Algorithms 2^［6］ 85.00 70.53
Algorithms 3^［21］ 91.30 —
Algorithms 4^［22］ 90.50 8.10
Ours 91.70 11.74

Tools

Get Citation

Copy Citation Text

Xiaochang Fan, Yu Liang, Wei Zhang. Infrared Vehicle Detection Algorithm Based on Improved Shuffle-RetinaNet[J]. Laser & Optoelectronics Progress, 2023, 60(24): 2410006

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites