Consistency-Aware Anchor Pyramid Network for Crowd Localization.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Crowd localization methods struggle with varying crowd densities and prediction inconsistencies.
• Existing fixed anchor strategies lead to suboptimal performance in crowded scenes.

Purpose of the Study:

• To develop an improved crowd localization method addressing anchor deficiency/excess and prediction inconsistencies.
• Introduce the Consistency-Aware Anchor Pyramid Network (CAAPN) for enhanced crowd density estimation.

Main Methods:

• Propose the Adaptive Anchor Generator (AAG) for density-adaptive anchor generation.
• Introduce the Localizer with Augmented Matching (LAM) to improve prediction consistency during training.
• Utilize a pyramid network structure for multi-scale feature extraction.

Main Results:

• CAAPN achieves state-of-the-art performance on five challenging crowd localization datasets.
• The Adaptive Anchor Generator effectively handles varying crowd densities.
• The Localizer with Augmented Matching reduces ranking inconsistencies, improving inference performance.

Conclusions:

• The proposed CAAPN significantly advances crowd localization accuracy and robustness.
• The adaptive anchor generation and augmented matching strategies are key to the method's success.
• This work provides a more effective solution for predicting human positions in crowded environments.