Table 1.
Summary of different methods from the latest literature.
Fig 1.
YOLOv5 network architecture.
Fig 2.
CA attention mechanism.
Fig 3.
CBAM attention mechanism.
Fig 4.
ECA attention mechanism.
Fig 5.
SE attention mechanism.
Fig 6.
Ghost module network structure.
Fig 7.
FPN, PAN, and BiFPN structures.
Fig 8.
Schematic diagram of EIOU-loss.
Fig 9.
(a) belt, (b) helmet and trunk, (c) incorrect helmet and vest, and (d) smoking and head.
Table 2.
Human Unsafe Behavior Dataset Distribution.
Fig 10.
Data enhancement.
Table 3.
Insecure state data set distribution of objects.
Fig 11.
Improved attention module.
Table 4.
Experimental results of the first improvement of the attention mechanism.
Table 5.
Experimental results of the second improvement of the attention mechanism.
Table 6.
Experimental results of the third improvement of attention mechanisms.
Fig 12.
Conv module replaced with GhostConv module.
Fig 13.
C3 module replaced with C3Ghost module.
Fig 14.
Mixing the first two improvements.
Table 7.
Ghost improvement experiment results.
Table 8.
Results of the first group of experiments.
Table 9.
Experimental results of the second group.
Table 10.
Results of the third group of experiments.
Fig 15.
The way the first group of experiments was improved.
Fig 16.
The way the second group of experiments was improved.
Fig 17.
The way the third group of experiments was improved.
Fig 18.
BiFPN (concat) improvement approach.
Table 11.
BiFPN improvement experimental results.
Table 12.
Experimental results of loss function improvement.
Table 13.
Results of NMS improvement experiments.
Table 14.
Results of ablation experiments.
Fig 19.
YOLO-CGBSE network architecture.
Fig 20.
P curve, R curve, PR curve and F1 curve.
Table 15.
Model comparison results.
Fig 21.
Comparison of detection results.
The first column of images is the image to be detected, the second column of images is the original YOLOv5n detection effect image, the third column of images is the YOLO-CGBSE detection effect image, Fig (a) is the UAV viewpoint image, Fig (b) is the high-resolution image, Fig (c) is the indoor operation image, Fig (d) is the high altitude operation image, and Fig (e) is the occluded high altitude operation image.
Table 16.
Object unsafe state model performance.
Fig 22.
Detection effect of unsafe state of object.
The first column of pictures is the picture to be detected, the second column of pictures is the original YOLOv5n detection effect, the third column is the YOLO-M detection effect, (a) is the picture of the safety net, (b) is the picture of the scaffolding perforation.
Fig 23.
System framework diagram.
Fig 24.
Detection system display.
Fig 25.
Identify the system’s detection results.