HA-CCN: Hierarchical Attention-based Crowd Counting Network.

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Single image-based crowd counting is crucial but challenging, especially in dense environments.
• Existing methods often struggle with accuracy in highly congested scenes.

Purpose of the Study:

• To develop an advanced crowd counting network that overcomes limitations in congested scenes.
• To introduce a novel weakly supervised approach for adapting crowd counting models to new datasets.

Main Methods:

• Proposed the Hierarchical Attention-based Crowd Counting Network (HA-CCN) utilizing VGG16.
• Incorporated a Spatial Attention Module (SAM) for low-level feature enhancement and Global Attention Modules (GAM) for high-level channel-wise information.
• Developed a weakly supervised framework using image-level labels for cross-dataset adaptation.

Main Results:

• Achieved state-of-the-art results on various crowd counting datasets.
• Demonstrated improved performance in highly congested scenes.
• Showcased effective adaptation to new datasets with reduced annotation effort.

Conclusions:

• HA-CCN offers a robust and efficient solution for single image crowd counting.
• The weakly supervised adaptation method significantly reduces annotation burden and enhances cross-dataset generalization.