HADF-Crowd: A Hierarchical Attention-Based Dense Feature Extraction Network for Single-Image Crowd Counting.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning

Background:

• Crowd counting is challenging due to variations in perspective, density, and scale.
• Existing Convolutional Neural Network (CNN) methods struggle with high perspective variations, leading to counting errors.
• Deep features in current CNNs are often used locally, hindering performance in complex scenes.

Purpose of the Study:

• To propose a novel CNN-based dense feature extraction network for accurate crowd counting.
• To address limitations in handling high perspective variations and feature propagation in existing methods.
• To improve counting accuracy by enhancing feature extraction and channel-wise attention.

Main Methods:

• Developed a CNN model with a backbone network, dense feature extraction modules (DFEMs), and a channel attention module (CAM).
• DFEMs utilize dense stacked convolution modules (DSCMs) for effective feature propagation from lower to higher layers.
• CAM is integrated to exploit class-specific responses and refine high-level features.

Main Results:

• The proposed method demonstrated improved performance on benchmark datasets (Shanghaitech Part-A, Part-B, and Venice).
• The dense connection strategy effectively propagated features, mitigating errors in high perspective-varying scenes.
• Channel attention enhanced the model's ability to distinguish foreground and background elements.

Conclusions:

• The proposed CNN-based dense feature extraction network offers a robust solution for accurate crowd counting.
• The integration of DFEMs and CAM effectively handles challenges posed by perspective variations and feature underestimation.
• The method shows competitive effectiveness compared to state-of-the-art crowd counting techniques.