Acta Photonica Sinica, Volume. 51, Issue 9, 0910001(2022)
Lightweight Real-time Detection Model of Infrared Pedestrian Embedded in Fine-scale
Figures & Tables(14)
Table 1. The allocation strategy of anchor
View tableView in ArticleTable 1. The allocation strategy of anchor
Detector size Target size Anchor size 16×16 Large target （38，132）
（62，205）
（116，321）
32×32 Medium target （13，49）
（19，60）
（25，89）
64×64 Small target （5，17）
（8，26）
（9，39）
Table 2. Hardware configuration
View tableView in ArticleTable 2. Hardware configuration
Equipment Model Quantity CPU Intel（R）Core（TM）i5-10400F 1 GPU NVIDIA GeForce RTX 2080Ti 1 RAM 8G DDR4 2666 2 Hard disk 512G SSD 1
Table 3. Parameter configuration
View tableView in ArticleTable 3. Parameter configuration
Parameter Value Learn rate 0.001 Epoch 600 Momentum 0.9 Hue 0.1 Policy Steps Steps 3 200，3 600
Table 4. Comparison of the effect of precision improvement strategies
View tableView in ArticleTable 4. Comparison of the effect of precision improvement strategies
Model Fine⁃scale detection layer 4 K⁃means++ Size/
MB
Speed/
（frame·s-1）
mAP/
%
Yolov4⁃tiny 23.5 90.7 80.6 Our √ 24.9 84.0 86.5 Our √ √ 24.9 79.9 87.5 Yolo⁃pedestrian √ √ √ 24.9 79.3 89.2
Table 5. Comparison of results of different fine-tuning strategies
View tableView in ArticleTable 5. Comparison of results of different fine-tuning strategies
Pruning rate Fine tuning Knowledge distillation Number of channels Size/MB mAP/% 0.00 2 912 24.9 89.2 0.75 √ 728 4.6 89.1 0.75 √ 728 4.6 89.5 0.80 √ 583 4.0 88.6 0.80 √ 583 4.0 89.2 0.85 √ 374 3.5 87.7 0.85 √ 374 3.5 88.0 0.90 √ 184 3.0 83.9 0.90 √ 184 3.0 84.1
Table 6. Comparison of different algorithms
View tableView in ArticleTable 6. Comparison of different algorithms
Model Size/MB Speed/（frame·s-1） mAP/% Yolov3 246.3 49.4 91.2 Yolov3-tiny 34.7 94.2 79.7 Yolov4 256 43.3 91.6 Yolov4-tiny 23.5 90.7 80.6 Yolo-pedestrian 24.9 79.3 89.2 TIPRD 4.0 88.7 89.2
Table 7. Comparison of detection speed of different models
View tableView in ArticleTable 7. Comparison of detection speed of different models
Model Speed/（frame·s-1） Yolov4 / Yolov3 0.95 Yolov4-tiny 5.2 Yolov3-tiny 2.6 TIPRD 6.9
Tools
Get Citation
Copy Citation Text
Yinhui ZHANG, Pengcheng ZHANG, Zifen HE, Sen WANG. Lightweight Real-time Detection Model of Infrared Pedestrian Embedded in Fine-scale[J]. Acta Photonica Sinica, 2022, 51(9): 0910001
Paper Information
Category:
Received: Mar. 4, 2022
Accepted: Apr. 21, 2022
Published Online: Oct. 26, 2022
The Author Email: HE Zifen (zyhhzf1998@163.com)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20