Accurate crowd counting in highly congested scenes is essential for public safety and effective resource management. The proposed method utilizes the VGG16 architecture enhanced with pretrained weights from EfficientNetB7 and compares its performance against the VGG16, ResNET50, MCNN, and VGG19 models. These models are trained and tested on the ShanghaiTech Part A and Part B datasets, representing densely and sparsely populated scenes under fog and rain, and bad weather conditions. The philosophy of this paper is based on achieving an optimal balance between enhancing image details and reducing noise to ensure high-quality feature extraction. Introducing a novel approach that uses various preprocessing techniques, including cubic interpolation and sharpening filters, to improve the quality and detail of images. After applying bad, harsh weather environments, such as rain and fog, the study analyzed the effect of preprocessing on five algorithms. The optimum results were achieved when concatenating EfficientNetB7 with VGG16 and using standalone VGG16. In contrast, VGG19 showed the poorest performance. This highlights the effectiveness of certain architectures under preprocessing enhancements. Combining the feature extraction power of <Emphasis Type="BoldItalic">EfficientNetB7</Emphasis> with <Emphasis Type="BoldItalic">VGG16</Emphasis> helps the model handle different crowd densities better. This makes the system more effective at analyzing crowds with varying numbers of people. Experiments show that the VGG16 model with EfficientNetB7 pretrained weights performs much better than the other five models in calculating error loss and the counting process. On the <Emphasis Type="BoldItalic">ShanghaiTech</Emphasis> Part <Emphasis Type="BoldItalic">A</Emphasis> and Part <Emphasis Type="BoldItalic">B</Emphasis> datasets, it achieved a mean absolute error (MAE) of 102.83 compared to an MAE of 142.26 for VGG16 and an MAE of 173.06 for VGG19. This shows that EfficientNetB7 helps VGG16 handle different crowd scales and densities in complex environments more effectively. This highlights the importance of VGG16 in extracting features and achieving strong performance in the crowd counting process, which can be used as a concatenated with a strong architecture. In addition, <Emphasis Type="BoldItalic">Mall</Emphasis> datasets scored the best metrics with a mean absolute error (MAE) of 0.78 and a mean square error (MSE) of 1.00549. Additionally, two datasets were used in testing, <Emphasis Type="BoldItalic">JHU-CROWD</Emphasis> + + + and <Emphasis Type="BoldItalic">UCF-QNRF</Emphasis>. The models were tested under different weather conditions, including rainy and foggy environments. Overall, the testing results were quite satisfactory.

Novel approach for crowd counting combining VGG16 and efficientnetb7 for optimal performance in harsh weather

期刊论文
发表日期：

2025年7月30日

查看全文

Heba F. Elsepae,
El-Sayed M. El-Rabaie,
Ehab K. I. Hamad,
Heba M. El-Hoseny

摘要

Accurate crowd counting in highly congested scenes is essential for public safety and effective resource management. The proposed method utilizes the VGG16 architecture enhanced with pretrained weights from EfficientNetB7 and compares its performance against the VGG16, ResNET50, MCNN, and VGG19 models. These models are trained and tested on the ShanghaiTech Part A and Part B datasets, representing densely and sparsely populated scenes under fog and rain, and bad weather conditions. The philosophy of this paper is based on achieving an optimal balance between enhancing image details and reducing noise to ensure high-quality feature extraction. Introducing a novel approach that uses various preprocessing techniques, including cubic interpolation and sharpening filters, to improve the quality and detail of images. After applying bad, harsh weather environments, such as rain and fog, the study analyzed the effect of preprocessing on five algorithms. The optimum results were achieved when concatenating EfficientNetB7 with VGG16 and using standalone VGG16. In contrast, VGG19 showed the poorest performance. This highlights the effectiveness of certain architectures under preprocessing enhancements. Combining the feature extraction power of EfficientNetB7 with VGG16 helps the model handle different crowd densities better. This makes the system more effective at analyzing crowds with varying numbers of people. Experiments show that the VGG16 model with EfficientNetB7 pretrained weights performs much better than the other five models in calculating error loss and the counting process. On the ShanghaiTech Part A and Part B datasets, it achieved a mean absolute error (MAE) of 102.83 compared to an MAE of 142.26 for VGG16 and an MAE of 173.06 for VGG19. This shows that EfficientNetB7 helps VGG16 handle different crowd scales and densities in complex environments more effectively. This highlights the importance of VGG16 in extracting features and achieving strong performance in the crowd counting process, which can be used as a concatenated with a strong architecture. In addition, Mall datasets scored the best metrics with a mean absolute error (MAE) of 0.78 and a mean square error (MSE) of 1.00549. Additionally, two datasets were used in testing, JHU-CROWD + + + and UCF-QNRF. The models were tested under different weather conditions, including rainy and foggy environments. Overall, the testing results were quite satisfactory.