Counting Crowd by Weighing Counts: A Sequential Decision-Making Perspective.

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Existing crowd counting models predominantly use one-step estimation.
• This approach often struggles with accuracy and generalization in complex crowd scenes.

Purpose of the Study:

• To propose a new crowd counting method by formulating it as a sequential decision-making (SDM) problem.
• To introduce LibraNet, a novel 'counting scale' agent inspired by human counting and physical weighing processes.

Main Methods:

• Decomposing crowd counting into sequential sub-decision problems, mirroring a scale weighing process.
• Implementing LibraNet as an agent that learns to place 'weights' (actions) based on image features and current state to balance a 'pointer' (estimated count).
• Exploring different state definitions and four implementation types: deep Q-network (DQN), actor-critic (AC), imitation learning (IL), and mixed AC+IL.

Main Results:

• LibraNet effectively mimics the physical scale weighing analogy for crowd counting.
• The proposed method achieves competitive or superior performance compared to state-of-the-art approaches across five benchmark datasets.
• LibraNet demonstrates significant cross-dataset generalization capabilities.

Conclusions:

• Crowd counting can be effectively modeled as a sequential decision-making problem.
• LibraNet offers a robust and generalizable solution for crowd counting, outperforming existing methods.
• LibraNet can serve as a plug-in module to enhance existing crowd counting models.