Meta-Knowledge and Multi-Task Learning-Based Multi-Scene Adaptive Crowd Counting.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Surveillance systems require crowd counting methods with strong generalization to unknown scenes.
• Diverse scenes necessitate scene-specific adaptation for optimal crowd counting performance.
• Balancing generalization and adaptation is a key challenge in crowd counting.

Purpose of the Study:

• To propose a multi-scene adaptive crowd counting method.
• To achieve strong generalization capability for unknown scenes.
• To effectively adapt to diverse scenes for improved accuracy.

Main Methods:

• A coarse-to-fine pipeline integrating a meta-knowledge network and multi-task learning.
• A generic two-stream network to encode meta-knowledge, including inter-frame temporal knowledge.
• A multi-task learning framework treating crowd density map regression as a homogeneous subtask for scene-specific parameter learning.

Main Results:

• The proposed method demonstrates improved accuracy compared to AMSNet and MAML-counting.
• Achieved a 10.29% reduction in Mean Absolute Error (MAE) compared to AMSNet.
• Achieved a 13.48% reduction in MAE compared to MAML-counting.

Conclusions:

• The method effectively balances generalization and scene-specific adaptation in crowd counting.
• Enables deployment to multiple new scenes without redundant model training.
• Offers a robust and accurate solution for adaptive crowd counting in real-world surveillance.