Optics and Precision Engineering, Volume. 32, Issue 24, 3616(2024)
Multi-feature cross UAV image detection algorithm under cross-layer attentional interaction
Figures & Tables(15)
Fig. 1. Multi-feature crossover UAV target detection network under cross-layer attentional interaction
Fig. 2. Sructure of the adaptive cross-layer attention interaction module
Fig. 4. Structure diagram of the multi-scale feature fusion module
Fig. 5. Model visualisation process for the MCAI network （features of interest in red）
Fig. 8. Visualization of the actual application scenario of the VisDrone2019-DET dataset
Fig. 9. Visualization of the actual application scenario of LZ Traffic Video
Table 1. Improved module ablation experiments
View tableView in ArticleTable 1. Improved module ablation experiments
Methods R P mAP0.5 mAP0.5∶0.95 mAP RT-DETR 40.8 56.5 42.5 25.7 34.2 G-ACAI 41.9 57.6 44.4 27.5 35.6 G-DMHSA 43 58.6 45.5 28.2 36.5 G-MSCF 44.2 58.8 46.1 28.5 36.8
Table 2. Category error experiments of the improved module
View tableView in ArticleTable 2. Category error experiments of the improved module
Methods Cls Loc Dupe Bkg Miss FP FN RT-DETR 20.98 3.82 0.21 2.12 13.31 10.92 37.54 G-ACAI 20.79 3.30 0.43 2.24 14.07 11.26 37.08 G-DMHSA 20.81 3.26 0.38 2.17 13.86 11.05 36.65 G-MSCF 20.25 3.29 0.37 2.27 13.76 11.12 36.70
Table 3. Robustness experiments on rainy weather on the VisDrone2019-DET dataset
View tableView in ArticleTable 3. Robustness experiments on rainy weather on the VisDrone2019-DET dataset
Methods mAP0.5@clean mAP0.5@rain mAP0.5∶0.95@clean mAP0.5∶0.95@rain YOLOv5-Decoder 33.9 21.5 19.9 11.7 YOLOv8-Deocder 32.0 20.9 18.3 11.5 RT-DETR 42.5 32.6 25.7 19 MCAI 46.1 34.7 28.5 20.7
Table 4. Comparison of results of different algorithmic models on VisDrone2019-DET
View tableView in ArticleTable 4. Comparison of results of different algorithmic models on VisDrone2019-DET
Method Backbone mAP mAP0.5 Param Faster-RCNN ResNet50 21.5 35.7 41.7 RetinaNet ResNet50 16.1 27.3 21.37 CenterNet DLA-34 12.4 22.70 25.6 SSD VGG-16 8.6 16.6 4.23 YOLOv3 MobileNet 8.1 17.7 3.75 YOLOv4 ResNet50 18.6 25.38 6.06 YOLOV5s CSPDarkNet 19.1 29.8 7.05 YOLOv7 CSPDarkNet 22.9 34.5 37.26 RT-DETR ResNet18 32.6 42.5 21.3 ATSS-FPN-DyHead ResNet50 20.4 33.8 38.91 TOOD ResNet50 20.4 33.9 32.4 DINO DETR 25.3 44.5 47.56 YOLOX-Tiny CSPDarkNet 14.8 27.8 5.035 GFL CSPDarkNet 19.3 32.1 32.279 RTMDet CSPDarkNet 18.4 31.2 4.876 MCAI ResNet18 36.8 46.1 22.4
Table 5. Overall performance comparison of BDD-100K dataset
View tableView in ArticleTable 5. Overall performance comparison of BDD-100K dataset
Method Year BackBone mAP0.5/% Faster-RCNN 2016 ResNet50 34.45 SSD 2017 VGG-16 28.42 RetinaNet 2018 ResNet50 37.65 CenterNet 2020 DLA-34 40.40 YOLOv3 2018 MobileNet 42.4 YOLOv4 2020 CSPDarkNet 45.3 YOLOv5s 2020 CSPDarkNet 61 YOLOv7 2022 CSPDarkNet 48.7 RTDETR 2023 ResNet18 70.2 ATSS-FPN-DyHead 2024 ResNet50 66.9 TOOD 2021 ResNet50 69.1 DINO 2023 ResNet50 70.4 YOLOX-Tiny 2021 CSPDarkNet 62.3 GFL 2020 CSPDarkNet 66.8 RTMDet 2023 CSPDarkNet 68.4 MCAI 2024 ResNet18 72.4
Table 6. Overall performance comparison of LZ traffic video detection dataset
View tableView in ArticleTable 6. Overall performance comparison of LZ traffic video detection dataset
Baseline Recall Precision mAP0.5 mAP0.5∶0.95 YOLOv3 41.4 82.6 52.1 30.9 YOLOv5 48.5 88.1 56.7 38.6 YOLOv6 40.4 72 45.2 31.7 YOLOv8 47.4 61.3 52.1 36.7 YOLOv8-
Decoder
49.5 76.7 55.1 37.2 RT-DETR 64.9 87.9 76.1 51.7 MCAI 75.8 82.8 80.6 59.9
Tools
Get Citation
Copy Citation Text
Zhihao ZHANG, Lixia DU, Yue HOU, Ziwei HAO, Jie YIN. Multi-feature cross UAV image detection algorithm under cross-layer attentional interaction[J]. Optics and Precision Engineering, 2024, 32(24): 3616
Paper Information
Category:
Received: Jun. 5, 2024
Accepted: --
Published Online: Mar. 11, 2025
The Author Email: HOU Yue (houyue@mail.lzjtu.cn)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20