Crowd Counting Based on Multiscale Spatial Guided Perception Aggregation Network.

Area of Science:

• Computer Vision
• Deep Learning
• Image Analysis

Background:

• Crowd counting using deep convolutional neural networks (CNNs) faces challenges like scale variation, non-uniform distribution, complex backgrounds, and occlusion.
• Existing CNN-based methods struggle to achieve high accuracy in crowded scenes due to these inherent difficulties.

Purpose of the Study:

• To propose an efficient and accurate crowd counting method overcoming limitations of existing approaches.
• To introduce the multiscale spatial guidance perception aggregation network (MGANet) for improved crowd density estimation.

Main Methods:

• Developed MGANet comprising a multiscale feature extraction network (MFEN), spatial guidance network (SGN), and attention fusion network (AFN).
• MFEN enhances scale adaptability; SGN captures spatial context and guides feature selection; AFN fuses features adaptively.
• Introduced a novel region-adaptive loss function to optimize error-prone regions and align training with evaluation metrics.

Main Results:

• MGANet demonstrated strong performance across challenging benchmarks: ShanghaiTech Part A/B, UCF-CC-50, UCF-QNRF, and JHU-CROWD++.
• Achieved superior recognition performance and robustness compared to state-of-the-art methods on multiple datasets.

Conclusions:

• The proposed MGANet effectively addresses scale variation, non-uniform distribution, and complex backgrounds in crowd counting.
• The region-adaptive loss function further enhances accuracy by focusing on difficult-to-recognize areas.