Infrared and Laser Engineering, Volume. 53, Issue 11, 20240256(2024)
An improved YOLOv8s method and its application in road traffic target detection
Figures & Tables(17)
Fig. 1. Effect comparison of visible images (a) and infrared images (b)
Fig. 6. Comparative illustration of the small object detection layer before and after its introduction
Fig. 7. Comparative diagram showing the effects before and after the introduction of the SPD module
Fig. 8. Test result diagram and thermal map effect comparison diagram I. (a) Original image; (b) Shallow heat map; (c) Intermediate heat map; (d) Deep heat map
Fig. 9. Test result diagram and thermal map effect comparison diagram II. (a) Original image; (b) Shallow heat map; (c) Intermediate heat map; (d) Deep heat map
Fig. 12. Comparison of test results. (a) YOLOv8s test result; (b) Test results for the improved network
Table 1. Different size prior box sizes
View tableView in ArticleTable 1. Different size prior box sizes
Feature map size Prior frame size [w/pixel, h/pixel] 20×20 [116,90] [156,198] [373,326] 40×40 [30,61] [62.45] [59,119] 80×80 [10,13] [16,30] [33,23] 160×160 [5,6] [8,14] [15,11]
Table 2. Results of ablation experiments with different improved methods
View tableView in ArticleTable 2. Results of ablation experiments with different improved methods
YOLOv8s New heads SPD My_att IOU P R mAP@0.5 mAP@0.5-0.95 Size/MB FPS/s √ 86.2% 75.1% 84.0% 55.5% 21.4 106.2 √ √ 86.6% 77.5% 86.8% 58.3% 21.9 94.2 √ √ √ 86.7% 79.6% 88.4% 59.7% 23.6 90.1 √ √ √ 86.8% 77.9% 87.5% 58.3% 22.4 86.9 √ √ √ 85.9% 77.7% 87.0% 57.8% 21.7 87.3 √ √ √ √ 86.7% 80.6% 88.9% 60.6% 24.5 85.8 √ √ √ √ √ 87.6% 80.6% 89.2% 60.7% 24.6 84.8
Table 3. Comparison of different attention mechanisms
View tableView in ArticleTable 3. Comparison of different attention mechanisms
Attention P R mAP@0.5 mAP@0.5-0.95 Size/MB FPS/s SE 86.8% 78.9% 88.0% 59.5% 24.0 87.5 CBAM 86.2% 79.3% 88.0% 59.2% 24.6 82.7 CA 86.7% 80.1% 88.4% 59.6% 25.0 84.0 EMA 87.1% 79.3% 88.3% 59.6% 25.2 79.3 My_att 86.7% 80.6% 88.9% 60.6% 24.5 85.8
Table 4. Comparison results of different loss functions
View tableView in ArticleTable 4. Comparison results of different loss functions
IOU loss mAP@0.5 mAP@0.5-0.95 FPS/s CIOU 88.9% 60.2% 85.8 F-CIOU 89.1% 60.4% 85.4 EIOU 89.1% 60.4% 84.7 F-EIOU 89.2% 60.7% 84.8
Table 5. Different target detection algorithms compare experimental results
View tableView in ArticleTable 5. Different target detection algorithms compare experimental results
Model P R mAP@0.5 mAP@0.5-0.95 Size/MB FPS/s YOLOv5s 85.0% 73.0% 82.7% 54.5% 18.5 99.4 YOLOv5m 87.2% 75.3% 84.9% 57.4% 50.5 79.5 YOLOv7 86.1% 78.6% 86.7% 54.7% 74.8 57.8 YOLOv8n 84.6% 69.4% 80.0% 51.3% 6.2 106.4 YOLOv8s 86.2% 73.5% 83.4% 55.5% 21.4 90.2 YOLOv8m 85.8% 75.9% 85.0% 57.8% 52.0 87.2 YOLOv8l 87.0% 76.2% 85.6% 59.0% 87.6 76.8 YOLOv8x 86.6% 76.8% 86.1% 59.5% 136.7 63.4 YOLOv9e 88.0% 77.7% 87.4% 61.4% 139.9 62.7 YOLOv9c 87.3% 76.2% 85.3% 58.4% 102.8 67.0 Ours 87.6% 80.6% 89.3% 60.7% 24.5 84.8
Tools
Get Citation
Copy Citation Text
Jiageng SANG, Zhijia ZHANG, Chuanmin XIAO, Haibo LUO, Junyao ZHANG. An improved YOLOv8s method and its application in road traffic target detection[J]. Infrared and Laser Engineering, 2024, 53(11): 20240256
Paper Information
Category: 图像处理
Received: Jun. 11, 2024
Accepted: --
Published Online: Dec. 13, 2024
The Author Email:
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20