Skip to Content
You are currently on the new version of our website. Access the old version .
ElectronicsElectronics
  • Article
  • Open Access

4 May 2023

17 Pages

DAAM-YOLOV5: A Helmet Detection Algorithm Combined with Dynamic Anchor Box and Attention Mechanism

,
,
,
and
1
School of Computer Science and Technology, Anhui University of Technology, Maanshan 243000, China
2
Research Institute of Information Technology, Anhui University of Technology, Maanshan 243000, China
*
Author to whom correspondence should be addressed.

Abstract

Helmet recognition algorithms based on deep learning aim to enable unmanned full-time detection and record violations such as failure to wear a helmet. However, in actual scenarios, weather and human factors can be complicated, which poses challenges for safety helmet detection. Camera shaking and head occlusion are common issues that can lead to inaccurate results and low availability. To address these practical problems, this paper proposes a novel helmet detection algorithm called DAAM-YOLOv5. The DAAM-YOLOv5 algorithm enriches the diversity of datasets under different weather conditions to improve the mAP of the model in corresponding scenarios by using Mosaic-9 data enhancement. Additionally, this paper introduces a novel dynamic anchor box mechanism, K-DAFS, into this algorithm and enhances the generation speed of the blocked target anchor boxes by using bidirectional feature fusion (BFF). Furthermore, by using an attention mechanism, this paper redistributes the weight of objects in a picture and appropriately reduces the model’s sensitivity to the edge information of occluded objects through pooling. This approach improves the model’s generalization ability, which aligns with practical application requirements. To evaluate the proposed algorithm, this paper adopts the region of interest (ROI) detection strategy and carries out experiments on specific, real datasets. Compared with traditional deep learning algorithms on the same datasets, our method effectively distinguishes helmet-wearing conditions even when head information is occluded and improves the detection speed of the model. Moreover, compared with the YOLOv5s algorithm, the proposed algorithm increases the mAP and FPS by 4.32% and 9 frames/s, respectively.

1. Introduction

Every year, the failure of workers to wear safety helmets results in significant losses for construction sites and society. In terms of monitoring the wearing of safety helmets, it is more common to use manual supervision. However, this method is unable to achieve real-time and all-weather monitoring. Solutions based on computer supervision offer advantages, such as fast response, full-time detection, and low monitoring cost, when compared with traditional monitoring methods. However, existing computer supervision technologies often fall short in terms of detection speed and accuracy, particularly in complex and open scenarios. In actual scenarios, areas close to the water, open fields, and strong winds can shake the camera during object detection, affecting the recognition rate and detection speed of the model. Changes in weather and lighting can also lead to difficulties in target recognition. Furthermore, the size of objects in images can differ significantly at varying distances, making it difficult to use fixed prior bounding boxes, resulting in inaccurate prior information and other issues. Additionally, during rush hours, large crowds and mutual occlusion due to people holding umbrellas on rainy days can increase missed-recognition rates. This paper aims to address the challenges of target occlusion and video frame shaking by proposing a helmet detection algorithm named Dynamic Anchor Box and Attention Mechanism YOLOv5 (DAAM-YOLOv5) that incorporates a dynamic anchor box and attention mechanisms. Our experimental results demonstrate that the proposed DAAM-YOLOv5 algorithm outperforms YOLOv5s in terms of both mAP (by 4.32%) and FPS (by 9 frames per second), indicating its superiority in real-time and all-weather helmet detection. This proves its effectiveness in improving worker safety at construction sites.
The main contributions of this paper are summarized as follows:
  • The improved Mosaic data enhancement method was used to randomly crop, scale, and arrange of up to nine (instead of four) pictures and make appropriate adjustments to light and angles to generate datasets with various brightness of light, angles of rotation, and target sizes. This method effectively increases the richness of the datasets and enhances the recognition ability of the algorithm under complex weather, light, shaking, and other conditions.
  • A new dynamic anchor box mechanism called Dynamic Anchor Feature Selection with K-means++ (K-DAFS) is proposed. Based on the dynamic anchor box [1], this mechanism adopts the K-means++ [2] clustering algorithm to automatically adjust the size of the training anchor boxes, accelerate the convergence of the target anchor boxes, solve the problem of mismatch between prior information and truth, and achieve faster detecting speed.
  • A new YOLOv5 model with attention mechanism is proposed. The attention mechanism was introduced to enhance the weight of the header information and weaken the background information. The ROI was pooled in the spatial attention module to properly adjust the sensitivity of the model to the object’s marginal information to improve the recognition rate of the model for the occlusion target.
  • It improves the GIoU in the original model structure, to CIoU, and solves the problem of slow convergence or non-convergence caused by GIoU_Loss degenerating to IoU when the prediction box and ground truth box are included.

3. DAAM-YOLOv5

3.1. Network Structure

The network structure of DAAM-YOLOv5 is shown in Figure 1. It is also divided into four parts, similar to YOLOv5: input, backbone, neck, and prediction.
Figure 1. Schematic diagram of the DAAM-YOLOv5 network structure model.
Where the “*” is multiplication sign, taking 320*320*32 in Figure 1 as an example, it means a photo with a length of 320 pixels, a width of 320 pixels, and a number of channels of 32.
Compared with YOLOv5, the main changes are as follows:
  • Mosaic-4 is modified to Mosaic-9 data enhancement in the Input;
  • A CBAM [5] attention mechanism is added to the Backbone;
  • A K-DAFS dynamic anchor box is added during the training phase and recognition phase;
  • In the Prediction part, CIoU is used to replace the original GIoU to compensate for the inherent defects of GIoU.
In this algorithm, the main functions of the input are Mosaic-9 data enhancement and the adaptive size of the anchor box. The original images generate new images through Mosaic-9, and then the newly generated images are sent to Backbone together with the original images. The backbone mainly contains the FOCUS structure, CBAM attention mechanism, and Cross Stage Partial_1 (CSP1_X) structure. The main function of the FOCUS structure is slicing the images. It splits the high-resolution feature maps into multiple low-resolution feature maps. The schematic diagram is shown in Figure 2. While reducing the number of parameters and network layers, FOCUS minimizes information loss and improves the convolution speed as a whole.
Figure 2. Slicing of Focus.
The CSP1_X structure can increase the gradient value of the backpropagation between layers to avoid the gradient disappearance caused by gradient descent, and so extract more fine-grained features without worrying about network degeneration [30]. The CBAM attention mechanism redistributes the recognition weight of the targets in the image and appropriately reduces the edge information sensitivity. The main function of prediction is to calculate and adjust the loss of generalized intersection over union (GIoU_Loss) and NMS and present the inference results. CIoU_Loss is used instead of GIoU_Loss to solve the problem that GIoU_Loss degenerates into IoU when the prediction box or the ground truth box is completely contained by the other boxes, which further relieves the matter that the convergence of IoU_Loss is slow. A diagram of the CIoU is shown in Figure 3, and the calculations of CIoU_Loss are defined in Equations (1)–(4).
C I o U = I o U − d C 2 − α ν ,
C I o U L O S S = 1 − C I o U ,
α = 0 , I o U < 0.5 ν 1 − I o U + ν , I o U ≥ 0.5 ,
ν = 4 π 2 tan − 1 ω g t h g t − tan − 1 ω p r e d h p r e d 2 ,
Figure 3. Diagram of CIoU.
In Figure 3, where d is the Euclidean distance between the center point of the prediction box and the ground truth box, c is the diagonal distance of the smallest circumscribed rectangle of the two boxes. In Equations (1)–(4), α is the weight coefficient and ν is the similarity of the aspect ratio between two boxes; the higher the similarity is, the better the prediction effect.

3.2. Improvements over the Original

3.2.1. Mosaic-9 Data Enhancement

As shown in the representation of Mosaic-9 data enhancement workflow in Figure 4, compared to Mosaic-4 in the original YOLOv5, this algorithm has made significant improvements in the random number and adjustment methods of images. First, in the Mosaic-9 process, a maximum of nine pictures are selected for cropping, zooming, arranging, and changing the brightness and then reintegrated and tiled to form a 640*640 picture; second, each of these pictures is rotated by plus or minus 10 degrees to form 18 rotated pictures, and then these 18 pictures are tiled and resized to 640*640. In the end, a total of 19 new dataset pictures were formed, which greatly enriched the type and quantity of the datasets.
Figure 4. Mosaic-9 data enhancement workflow.

3.2.2. K-DAFS Dynamic Anchor Box

In the one-stage target detection to which YOLO belongs, an anchor box is generally used instead of the target selection stage in two-stage object detection. Generally, there are three methods for selecting an anchor box: generating an anchor box through K-means [31] clustering, artificial regulation, and training. However, these three methods all have certain shortcomings. The K-means method is easily affected by outliers and initial values due to its limitations. Although manually setting anchor boxes is relatively flexible, it requires a lot of effort to find suitable anchor boxes, which increases the difficulty of training [32]. Training to generate anchor boxes is a commonly used method, but traditional training methods are from the top down, cannot be dynamically adjusted, and easily produce a prior information mismatch truth. To address these issues, this paper proposes a novel method for a dynamic anchor box, called K-DAFS, that combines dynamic anchor feature selection (DAFS) [1] with the K-means++ algorithm [2]. The module is shown in Figure 5. This module is constructed based on the anchor refinement module (ARM) [33]. For any nonterminal feature map, the relevant data information about the anchor box comes from both the upper and lower layers simultaneously.
Figure 5. DAFS model structure.
DAAM-YOLOv5 uses manually specified anchor boxes in the initial training, uses the K-means++ clustering algorithm to cluster the marked target anchor boxes multiple times, and generates multiple prior bounding boxes of different sizes. It filters the anchor boxes, which have an IoU of less than 0.5 with the ground truth box, through ARM. However, ARM will result in a mismatch between the receptive field of the point and the anchor boxes for some feature point, which may weaken the detection ability of the model [1]. Therefore, this paper fuses top-down and bottom-up bidirectional paths using bidirectional feature fusion (BFF) combined with K-means++ to connect feature maps of different scales and receive information from the upper and lower layers to dynamically adjust the anchor boxes. Finally, the dynamic refining anchor box is obtained as the prior bounding box for model training.
In detail, the workflow of K-means++ in BFF is shown in Figure 6, and is as follows:
Figure 6. Calculation diagram of K-means++ in BFF.
  • Among the prediction boxes, the set B is obtained through the feedback from the upper and lower levels of BFF, the center point of a prediction box is randomly selected as the initial center point u 0 .
  • The Euclidean distance d i ( x i , u 0 ) between u 0 and the four anchor points x i of the ground truth box is calculated, and the farthest point x t is obtained as the new center point u 1 .
    d i ( x i , u 0 ) = | | x i − u 0 | | 2 2 ,   i ϵ { 1 , 2 , 3 , 4 } ,
    d t ( x t , u 0 ) = M a x   d i ( x i , u 0 ) ,   t ϵ { 1 , 2 , 3 , 4 } ,
  • The minimum value E of the Euclidean distance between u 1 and the anchor point x i in all prediction boxes in B is calculated, then the anchor box j corresponding to E is the optimal anchor box.
    d j = ∑ i = 1 4 | | x i − u 1 | | 2 2 ,   j ϵ B ,
    E = Min   d j ,

3.2.3. Attention Mechanism

To improve the recognition accuracy of the algorithm in a specific area, a currently more advanced attention mechanism, CBAM [5], is introduced based on YOLOv5. The main method is to redistribute the weights of information and filter appropriate information to fit the current scene. Compared with the attention mechanism of SENet [6], which only focuses on channel information, CBAM can take into account the reception of spatial information while paying attention to channel information. Without obviously increasing the number of convolution operation layers, it is more effective for solving the problems of small edge areas, blurred edges, and partial information loss. Experiments have proven that [5], in the process of training and learning, channel attention is first used in sequence, and then spatial attention is used to achieve the best results. CBAM has a better performance in lightweight models, but it increases the complexity of the model to a certain extent and may reduce the detection speed by a small amount. The structural model of the CBAM attention mechanism is shown in Figure 7:
Figure 7. The structural model of the CBAM attention mechanism.
As shown in Figure 7, CBAM mainly includes two modules, namely, the channel attention module and the spatial attention module. The two focus on different objects. The former mainly focuses on the category of the current convolution object, and the latter focuses on the position of the current object. The former half of CBAM is the channel attention module, which first performs maximum pooling and average pooling on the input feature map and then roughly perceives the object category using the shared multilayer perceptron. Finally, sum up the data and then feed it into the activation function for processing after elementwise operation. The calculation process can be briefly described as Equations (9) and (10):
M c F = σ M L P A v g P o o l F + M L P M a x P o o l F ,
= σ W 1 W 0 F a v g C + W 1 W 0 F m a x C ,
where σ is the sigmoid function, C is the number of channels, and W 0 and W 1 are the weights of the first layer and the second layer of the perceptron, respectively.
In CBAM, the position information of the target is supplemented by introducing the spatial attention module. The pooling operation along the channel axis can effectively highlight the information area [34]. First, the output feature map in the attention channel is weighted with the original input feature map. The obtained image is pooled with the maximum and the average. Second, a 7*7 convolution kernel is used to convolve the two feature layers at the same time, and the two feature layers are aggregated into one feature layer through pooling. Finally, the feature after the layer is processed by the sigmoid function, M s , that contains the emphasized or suppressed position information that has been obtained. The calculation process for spatial attention can be briefly described as Equations (11) and (12):
M s F = σ f 7 × 7 ( [ A v g P o o l F ; M a x P o o l F ] ,
= σ f 7 × 7 F A v g C ; F M a x C ,
where σ is a sigmoid function, f 7 × 7 is a convolution operation with a size of 7 × 7, and F A v g C and F M a x C are the feature layers after average pooling and maximum pooling, respectively.

4. Experimental Results and Analysis

4.1. Experimental Datasets

The dataset of 32,000 images used in this paper was collected from specific scenes or generated by Mosaic-9 according to real images and includes 12,000 actually collected images and 20,000 automatically generated images. Considering the influence of the natural environment on the recognition rate, the images include sunny, cloudy, rainy, day, night, and other conditions. There were 123,688 positive samples (helmet), 75,289 negative samples (head), 256,832 human contour samples (people), and 57,855 unlabeled samples. The unlabeled samples had insufficient clarity, serious occlusion, or poor contour information, so they were not used for training. Finally, the mixed datasets were divided into training datasets, validation datasets and testing datasets, at a ratio of 7:2:1, for the experiments. The partial images generated by Mosaic-9 are shown in Figure 8, and some low-brightness images have been automatically generated to compensate for the lack of nighttime training datasets.
Figure 8. Some of the images generated by Mosaic 9.

4.2. Experimental Environments

Both the training and testing of this experiment were carried out using the Windows 10 operating system. The specific software and hardware environments are listed in Table 1.
Table 1. Training Environmental Configuration.

4.3. Experimental Hyperparation and Evaluation Criteria

To ensure the effectiveness of the experimental data, the same hyperparameters were used for training and testing different algorithm models. Stochastic gradient descent [35] backpropagation was used to fine-tune and optimize the network parameters, and the parameter settings are shown in Table 2. The initial learning rate was set to 0.01, weight decay was set to 0.00002, batch size was set to 32, IoU was set to 0.35, the momentum factor was set to 0.937, to avoid the model training falling into local optimum, and the number of epochs was set to 300.
Table 2. Environmental Hyperparameter Configuration.
Model evaluation is an important task in deep learning. The evaluation of indicators usually includes accuracy, precision, recall, mAP, parameters, and giga-floating-point operations per second (GFLOPs).
These are described by the following equations:
A c c u r a c y = T P + F N T P + T N + F N + F P ,
P r e c i s i o n = T P T P + F P ,
R e c a l l = T P T P + T N ,
where TP indicates the number of samples in which positive class samples are correctly predicted as positive, TN indicates the number of samples in which negative class samples are correctly predicted as negative, FP indicates the number of negative samples incorrectly predicted as positive samples and FN represents the number of positive samples that are wrongly predicted as negative samples.
The more widely used criterion in model evaluation is mAP, and the mAP calculation is as Equation (16):
m A P = 1 K ∑ i = 1 N A P i ;   A P = ∑ i = 1 n − 1 r i + 1 − r i P i n t e r r i + 1 ,
where r is the recall value corresponding to each interpolation segment of Precision sorted in ascending order.

4.4. Training Results

Figure 9 shows the comparison of mAP under the same configurations of the diverse algorithm models involved in this experiment for approximately 300 epochs. The results show that the mAP of each algorithm rapidly rises in the top 20 epochs, among which DAAM-YOLOv5 and YOLOv3 are the fastest rising; this paper’s proposed algorithm reached a mAP of 0.91 and YOLOv3s reached 0.83. The proposed algorithm converges stably at epoch = 52, which is later than the other algorithms. All algorithms tend to be stable at approximately epoch = 150, and the subsequent training makes the mAP slightly decrease, showing signs of overfitting. There is no significant difference between the CBAM and SE attention mechanisms in the first 40 epochs, and the mAP of CBAM is slightly higher than that of the SE attention mechanism after 40 epochs. Finally, the proposed algorithm’s optimal mAP is 0.9636, which is higher than the 0.9204 of the YOLOv5 model and increased by approximately 4.32%, verifying the feasibility of this improved algorithm.
Figure 9. The mAP of related algorithms in training.
The key data points for these algorithms are shown in Table 3.
Table 3. The mAP comparison of different algorithms in different epochs.

4.5. Detection Results of DAAM-YOLOv5 in Various Cases

Figure 10 shows the partial detection results of DAAM-YOLOv5 in multiple cases in the testing datasets. Among them, (a)–(c) show the situations in which a tester rides an electric bicycle through the detection area at a speed of 40 km/h. The algorithm can still correctly identify the targets in the case of high speed. Figure 10d shows the detection of ROI, the people are detected within the ROI, but not outside it. It can be seen in (e) that a safety helmet will not be detected in the case of no human wearing a safety helmet, because the identification of the safety helmet is not set when no one is wearing it, and the algorithm does not appear to be misidentified. In (f), the violator intentionally obscured the head information in an attempt to interfere with the detection results, but the algorithm still detected it correctly, indicating that the generalization of this algorithm is in good shape. For irregular safety helmets that do not meet the requirements of production, worn by the personnel in (g,h), the algorithm correctly detects the violation information. In (i), there is the case of insufficient light at night, and the detection effect is good.
Figure 10. Detection results of DAAM-YOLOv5 in multiple cases.

4.6. Comparison of Detection Effects

4.6.1. Comparison between the YOLOv5 Algorithm and DAAM-YOLOv5 Algorithm

Figure 11 compares the detection effects of YOLOv5 and DAAM-YOLOv5 under the same configuration. It is obvious that the improved DAAM-YOLOv5 has good performance in the case of occluded objects after adding the attention mechanism. It effectively reduces the false recognition and misidentification rates of the targets.
Figure 11. Comparison between YOLOv5 algorithm and DAAM-YOLOv5 algorithm.

4.6.2. Ablation Experiment Comparison

Figure 12 shows a detailed mAP of the algorithms in Figure 9 with epochs between 200 and 290. An ablation experiment was also performed to verify the impact of each module on the model performance. The results are shown in Table 4. The first row in the table is the YOLOv5 algorithm without improvement. In the original algorithm, Mosaic-4 data enhancement is used by default. Mosaic-4 data enhancement is used by default in the absence of Mosaic-9 data enhancement. The second and third rows show the comparison of the YOLOv5+CBAM and YOLOv5+SE attention mechanisms, respectively. The fourth row is the DAAM-YOLOv5 algorithm in this paper.
Figure 12. The mAP details of related algorithms with epochs from 0 to 75 and 100 to 299.
Table 4. Ablation experiment.
The results showed the following:
  • After the addition of CBAM alone, the mAP of the model has been improved by approximately 2.36%. However, compared with the YOLOv5 algorithm, the detection speed and recall have decreased to a certain extent, with a reduction of 17 frames/s and 2.81%, respectively. This indicates that the CBAM module can increase the mAP of the model, but it will slightly reduce the model detection speed and increase the missed-recognition rate.
  • Compared with the second and fourth lines, it can be seen that the precision of the SE attention mechanism is slightly lower than that of the CBAM attention mechanism, by 0.41%, but the detection speed is faster than that of CBAM attention mechanism by 9 frames/s.
  • Compared with the first and fourth lines, it can be seen that, after improving the Mosaic-9 data enhancement, the algorithm’s mAP has increased by 1.62%, while the speed rate has decreased by 11 frames/s. In addition, this module has shown significant improvements in both accuracy and recall, with gains of 4.35% and 8.68%, respectively.
  • According to the fifth line, it can be seen that there is a slight decrease in mAP, by approximately 0.96%, after adding the K-DAFS dynamic anchor box. However, the obvious improvements in the detection speed and recall of the algorithm are 23 frames/s and 10.97%, respectively. This indicates that the K-DAFS module can significantly increase the detection speed of the model while reducing the model’s missed-recognition rate.
  • After integrating the CBAM attention mechanism, the Mosaic-9 data augmentation, and the K-DAFS, the model’s detection speed was only slower than the K-DAFS model but it had the highest mAP. Compared with the original YOLOv5, the mAP, detection speed, and recall were improved by 4.32%, 9 frames/s, and 6.85%, respectively.

5. Conclusions

The impact of safety production on businesses and society is significant. However, existing safety helmet detection algorithms have difficulty meeting the needs of complex open scenarios. This paper proposes a safety helmet detection algorithm, DAAM-YOLOv5, which combines a dynamic anchor box with an attention mechanism and demonstrates superior performance in addressing significant target variations, target occlusions, and video frame jitter. The algorithm improves the model’s accuracy when dealing with significant variations in targets by adjusting the Mosaic data enhancement in the input layer of YOLOv5. In addition, a new dynamic anchor box is added, and BFF is used to dynamically adjust anchor box sizes to improve the model’s detection speed. By incorporating an attention mechanism to adjust the weighting of the background and target recognition, as well as the sensitivity of the algorithm to edge information, the issue of low recognition accuracy in occluded scenarios is resolved. Finally, the algorithm improves the original model structure by replacing GIoU with CIoU, addressing the problem of slow or failed convergence when the prediction box contains the ground truth box.
The results of experimental analysis show that the DAAM-YOLOv5 algorithm has improved mAP by 4.32% and detection speed by 9 FPS. In comparison to other object detection algorithms, we achieved higher accuracy and speed. Therefore, the improved algorithm is better suited for scenarios such as ports, smart construction sites, and danger warning zone. However, the algorithm’s recognition performance on flooded road surfaces is not significant at present. In the future, we will increase our research in this field to improve the recognition rate in this scenario.

Author Contributions

Conceptualization, W.T.; Methodology, W.T. and Z.W.; Software, X.H.; Validation, Z.W.; Formal analysis, X.H.; Investigation, J.C. and X.H.; Resources, Z.W. and J.C.; Data curation, W.L.; Writing—original draft, Z.W.; Visualization, W.T. and W.L. All authors have read and agreed to the published version of the manuscript.

Funding

This Paper was funded by Anhui Provincial Key Research and Development Project (Grant No. 202004a07020028); and Natural Science Foundation of Anhui Province, China (Grant No. 2008085QF305).

Data Availability Statement

Not applicable.

Conflicts of Interest

The authors declare no conflict of interest.

Abbreviations

The following abbreviations are used in this manuscript:
YOLOYou Only Look Once
BFFBidirectional Feature Fusion
ROIRegion of Interest
mAPMean Average Precision
DAFSDynamic Anchor Feature Selection
IoUIntersection of Union
CIoUComplete Intersection over Union
GIoUGeneralized Intersection over Union
SAMStereo Attention Module
CAMClass Activation Mapping
CBAMConvolutional Block Attention Module
SESqueeze-and-Excitation
R-CNNRegion based Convolutional Neural Network
R-FCNRegion-based Fully Convolutional Network
SSDSingle-shot MultiBox Detector
R-SSDRainbow Single-shot MultiBox Detector
DSSDDeconvolutional Single-shot MultiBox Detector
FSSDFeature Fusion Single-shot MultiBox Detector
NMSNon-maximum Suppression
CSP1_XCross Stage Partial_1
ARMAnchor Refinement Module
SGDStochastic Gradient Descent
GFLOPsGiga-floating-point Operations per Second

References

  1. Li, S.; Yang, L.; Huang, J.; Hua, X.S.; Zhang, L. Dynamic anchor feature selection for single-shot object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, Seoul, Republic of Korea, 27 October–2 November 2019; pp. 6609–6618. [Google Scholar]
  2. Arthur, D.; Vassilvitskii, S. k-Means++: The Advantages of Careful Seeding; Stanford University: Stanford, CA, USA, 2006. [Google Scholar]
  3. Zhou, B.; Khosla, A.; Lapedriza, A.; Oliva, A.; Torralba, A. Learning deep features for discriminative localization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Las Vegas, NV, USA, 27–30 June 2016; pp. 2921–2929. [Google Scholar]
  4. Ying, X.; Wang, Y.; Wang, L.; Sheng, W.; An, W.; Guo, Y. A stereo attention module for stereo image super-resolution. IEEE Signal Process. Lett. 2020, 27, 496–500. [Google Scholar] [CrossRef] [Scilit]
  5. Woo, S.; Park, J.; Lee, J.Y.; Kweon, I.S. Cbam: Convolutional block attention module. In Proceedings of the European Conference on Computer Vision (ECCV), Munich, Germany, 8–14 September 2018; pp. 3–19. [Google Scholar]
  6. Hu, J.; Shen, L.; Sun, G. Squeeze-and-excitation networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Salt Lake City, UT, USA, 18–23 June 2018; pp. 7132–7141. [Google Scholar]
  7. Girshick, R.; Donahue, J.; Darrell, T.; Malik, J. Rich feature hierarchies for accurate object detection and semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Columbus, OH, USA, 23–28 June 2014; pp. 580–587. [Google Scholar]
  8. Girshick, R. Fast r-cnn. In Proceedings of the IEEE International Conference on Computer Vision, Santiago, Chile, 7–13 December 2015; pp. 1440–1448. [Google Scholar]
  9. Ren, S.; He, K.; Girshick, R.; Sun, J. Faster r-cnn: Towards real-time object detection with region proposal networks. Adv. Neural Inf. Process. Syst. 2015, 28. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  10. He, K.; Gkioxari, G.; Dollár, P.; Girshick, R. Mask r-cnn. In Proceedings of the IEEE International Conference on Computer Vision, Venice, Italy, 22–29 October 2017; pp. 2961–2969. [Google Scholar]
  11. Dai, J.; Li, Y.; He, K.; Sun, J. R-fcn: Object detection via region-based fully convolutional networks. Adv. Neural Inf. Process. Syst. 2016, 29, 379–387. [Google Scholar]
  12. Redmon, J.; Divvala, S.; Girshick, R.; Farhadi, A. You only look once: Unified, real-time object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Las Vegas, NV, USA, 27–30 June 2016; pp. 779–788. [Google Scholar]
  13. Redmon, J.; Farhadi, A. YOLO9000: Better, faster, stronger. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, HI, USA, 21–26 July 2017; pp. 7263–7271. [Google Scholar]
  14. Redmon, J.; Farhadi, A. Yolov3: An incremental improvement. arXiv 2018, arXiv:1804.02767. [Google Scholar]
  15. Bochkovskiy, A.; Wang, C.Y.; Liao, H.Y.M. Yolov4: Optimal speed and accuracy of object detection. arXiv 2020, arXiv:2004.10934. [Google Scholar]
  16. Liu, W.; Anguelov, D.; Erhan, D.; Szegedy, C.; Reed, S.; Fu, C.Y.; Berg, A.C. Ssd: Single shot multibox detector. In Proceedings of the Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, 11–14 October 2016; Proceedings, Part I 14. Springer International Publishing: Amsterdam, The Netherlands, 2016; pp. 21–37. [Google Scholar]
  17. Jeong, J.; Park, H.; Kwak, N. Enhancement of SSD by concatenating feature maps for object detection. arXiv 2017, arXiv:1705.09587. [Google Scholar]
  18. Fu, C.Y.; Liu, W.; Ranga, A.; Tyagi, A.; Berg, A.C. Dssd: Deconvolutional single shot detector. arXiv 2017, arXiv:1701.06659. [Google Scholar]
  19. Li, Z.; Zhou, F. FSSD: Feature fusion single shot multibox detector. arXiv 2017, arXiv:1712.00960. [Google Scholar]
  20. Lin, T.Y.; Goyal, P.; Girshick, R.; He, K.; Dollár, P. Focal loss for dense object detection. In Proceedings of the IEEE International Conference on Computer Vision, Venice, Italy, 22–29 October 2017; pp. 2980–2988. [Google Scholar]
  21. Han, K.; Zeng, X. Deep Learning-Based Workers Safety Helmet Wearing Detection on Construction Sites Using Multi-Scale Features. IEEE Access 2022, 10, 718–729. [Google Scholar] [CrossRef] [Scilit]
  22. Li, M.; Zhang, W.; Hu, B.; Kang, J.; Wang, Y.; Lu, S. Automatic Assessment of Depression and Anxiety through Encoding Pupil-wave from HCI in VR Scenes. ACM Trans. Multimid. Comput. Commun. Appl. 2022. [Google Scholar] [CrossRef] [Scilit]
  23. Yan, J.; Zhao, L.; Diao, W.; Wang, H.; Sun, X. AF-EMS Detector: Improve the Multi-Scale Detection Performance of the Anchor-Free Detector. Remote Sens. 2021, 13, 160. [Google Scholar] [CrossRef] [Scilit]
  24. Tan, S.; Lu, G.; Jiang, Z.; Huang, L. Improved YOLOv5 network model and application in safety helmet detection. In Proceedings of the 2021 IEEE International Conference on Intelligence and Safety for Robotics (ISR), Nagoya, Japan, 4–6 March 2021; pp. 330–333. [Google Scholar]
  25. Li, J.; Liu, H.; Wang, T.; Jiang, M.; Wang, S.; Li, K.; Zhao, X. Safety helmet wearing detection based on image processing and machine learning. In Proceedings of the 2017 Ninth International Conference on Advanced Computational Intelligence (ICACI), Doha, Qatar, 4–6 February 2017; pp. 201–205. [Google Scholar]
  26. Qing, T.; Yuan, H.; Fei, D.; Yao, N. Research on Head and Shoulders Detection Algorithm in Complex Scene Based on YOLOv5. In Proceedings of the 2022 IEEE Asia-Pacific Conference on Image Processing, Electronics and Computers (IPEC), Dalian, China, 14–16 April 2022; pp. 368–371. [Google Scholar]
  27. Dai, J.; Qi, H.; Xiong, Y.; Li, Y.; Zhang, G.; Hu, H.; Wei, Y. Deformable convolutional networks. In Proceedings of the IEEE International Conference on Computer Vision, Venice, Italy, 22–29 October 2017; pp. 764–773. [Google Scholar]
  28. Nan, X.; Lu, W.; Chongliu, J.; Yuemou, J.; Xiaoxia, M. Simulation of Occluded Pedestrian Detection Based on Improved YOLO [J/OL]. J. Syst. Simul. 2023, 35, 286. [Google Scholar] [CrossRef]
  29. Dai, B.; Nie, Y.; Cui, W.; Liu, R.; Zheng, Z. Real-time Safety Helmet Detection System based on Improved SSD. In Proceedings of the 2nd International Conference on Artificial Intelligence and Advanced Manufacture, Manchester, UK, 15–17 October 2020; pp. 95–99. [Google Scholar]
  30. Orhan, A.E.; Pitkow, X. Skip connections eliminate singularities. arXiv 2017, arXiv:1701.09175. [Google Scholar]
  31. Tibshirani, R.; Walther, G.; Hastie, T. Estimating the number of clusters in a data set via the gap statistic. J. R. Stat. Soc. Ser. B (Stat. Methodol.) 2001, 63, 411–423. [Google Scholar] [CrossRef] [Scilit]
  32. Zhou, X.; Lan, X.; Zhang, H.; Tian, Z.; Zhang, Y.; Zheng, N. Fully convolutional grasp detection network with oriented anchor box. In Proceedings of the 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), Madrid, Spain, 1–5 October 2018; pp. 7223–7230. [Google Scholar]
  33. Sloss, A.; Symes, D.; Wright, C. ARM System Developer’s Guide: Designing and Optimizing System Software; Elsevier: Amsterdam, The Netherlands, 2004. [Google Scholar]
  34. Zagoruyko, S.; Komodakis, N. Paying more attention to attention: Improving the performance of convolutional neural networks via attention transfer. arXiv 2016, arXiv:1612.03928. [Google Scholar]
  35. Song, S.; Chaudhuri, K.; Sarwate, A.D. Stochastic gradient descent with differentially private updates. In Proceedings of the 2013 IEEE Global Conference on Signal and Information Processing, Austin, TX, USA, 3–5 December 2013; pp. 245–248. [Google Scholar]
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Article Metrics

Citations

Article Access Statistics

Multiple requests from the same IP address are counted as one view.