A Crowded Object Counting System with Self-Attention Mechanism

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Traditional object detection struggles with counting small, crowded objects, leading to inaccuracies.
• Existing density map estimation models face challenges in achieving high accuracy for dense object counting.

Purpose of the Study:

• To develop a novel crowded object counting system using density map estimation.
• To enhance the accuracy of density map generation for improved object counting in dense scenes.

Main Methods:

• Proposed a novel model integrating a context-aware network with a self-attention mechanism for density map estimation.
• Relabeled the TRANCOS database to provide more complete ground truth data.
• Analyzed the parameters of the self-attention mechanism to find optimal combinations.

Main Results:

• Achieved high accuracy rates on multiple benchmark datasets: TRANCOS (85.9%), relabeled TRANCOS (90.0%), ShanghaiTech Part A (83.4%), and Part B (92.6%).
• Demonstrated the effectiveness of the self-attention mechanism in improving density map estimation accuracy.
• An ablation study confirmed the optimal parameter combination for the context-aware network with self-attention.

Conclusions:

• The proposed model significantly enhances crowded object counting accuracy compared to existing methods.
• Data augmentation through relabeling and architectural improvements like self-attention are crucial for dense counting tasks.
• The study provides a robust framework for accurate object counting in challenging, crowded environments.