A Hybrid Multi-Scale Transformer-CNN UNet for Crowd Counting.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Crowd counting is vital for public security and smart cities.
• Existing deep learning models struggle with scale variation, occlusion, and clutter.
• Advanced crowd density estimation is crucial for effective crowd management.

Purpose of the Study:

• To develop a novel deep learning network for accurate crowd counting.
• To address challenges like extreme scale variations and severe occlusion.
• To enhance crowd density estimation in complex scenarios.

Main Methods:

• Proposed a Hybrid Multi-Scale Transformer-CNN U-shaped Network (HMSTUNet).
• Integrated a Multi-Scale Vision Transformer (MSViT) for long-range dependencies.
• Utilized a Dynamic Convolutional Attention Block (DCAB) for local density patterns.
• Employed a U-shaped encoder-decoder with skip connections for feature fusion.

Main Results:

• HMSTUNet achieved state-of-the-art performance on five public benchmarks.
• The model attained the best Mean Absolute Error (MAE) on all datasets.
• Achieved the best Mean Squared Error (MSE) on three out of five datasets.
• Demonstrated superior robustness and generalization capabilities.

Conclusions:

• HMSTUNet effectively addresses key challenges in crowd counting.
• The proposed hybrid architecture offers significant improvements over existing methods.
• The model shows strong potential for real-world applications in public security and smart cities.