SPCANet: congested crowd counting via strip pooling combined attention network.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning

Background:

• Crowd counting is crucial for public safety and management, but current models struggle with perspective distortion, occlusion, and irregular crowd distribution in dense scenes.
• Inaccurate spatial information capture limits the effectiveness of existing crowd counting technologies.

Purpose of the Study:

• To develop an advanced crowd counting model, SPCANet, that overcomes limitations in handling highly crowded and noisy environments.
• To improve the accuracy and robustness of crowd density estimation by addressing issues like perspective distortion and occlusion.

Main Methods:

• Proposed a novel network model, SPCANet, integrating normed-deformable convolution (NDConv) with strip pooling and efficient channel attention (ECA).
• Strip pooling utilizes long, narrow kernels (1xN or Nx1) to effectively model long-distance dependencies and handle dense, occluded crowds.
• Efficient channel attention (ECA) employs local cross-channel interactions and 1D convolution to enhance model performance with reduced complexity.

Main Results:

• SPCANet achieved state-of-the-art performance on four benchmark datasets (Shanghai Tech Part A & B, UCF-QNRF, UCF CC 50), outperforming baseline models.
• Demonstrated significant improvements in mean absolute error (MAE) and a 5.7% average decrease in mean squared error (MSE) across datasets.
• The model exhibited greatly improved robustness in estimating crowd counts under challenging conditions.

Conclusions:

• SPCANet effectively addresses the challenges of crowd counting in dense, occluded, and perspective-distorted scenes.
• The integration of strip pooling and efficient channel attention offers a promising direction for advancing crowd counting technology.
• The proposed method shows significant practical applicability in real-world crowd management and analysis scenarios.