Deformable Density Estimation via Adaptive Representation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Crowd counting is crucial for public safety and increasingly utilizes convolutional neural networks (CNNs).
• Existing CNN-based methods generate density maps but struggle with significant target scale variations due to perspective effects.
• This limitation hinders accurate crowd counting in scenes with diverse target sizes.

Purpose of the Study:

• To develop a scale-sensitive crowd density map estimation framework to overcome the limitations of existing methods.
• To improve crowd counting performance by effectively handling target scale variations.
• To enhance the accuracy of crowd analysis in public safety applications.

Main Methods:

• Proposed a scale-sensitive crowd density map estimation framework incorporating Adaptive Density Map (ADM) and Deformable Density Map Decoder (DDMD).
• ADM generates scale-aware density maps by adaptively varying Gaussian kernel sizes based on target size.
• DDMD employs deformable convolutions to better fit Gaussian kernel variations and enhance scale sensitivity, guided by an Auxiliary Branch during training.

Main Results:

• Experiments on large-scale datasets demonstrate the effectiveness of the proposed ADM and DDMD.
• Visualizations confirm that deformable convolutions successfully learn and adapt to target scale variations.
• The framework significantly improves crowd counting performance compared to existing methods.

Conclusions:

• The proposed scale-sensitive framework effectively addresses the challenge of target scale variation in crowd counting.
• ADM and DDMD are key components that enhance density map generation and network sensitivity to scale changes.
• This work offers a significant advancement in crowd analysis for public safety and related fields.