Counting People by Estimating People Flows.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Current crowd counting methods primarily use deep networks on individual images.
• Few methods leverage temporal consistency in videos, and existing ones use weak smoothness constraints.

Purpose of the Study:

• To develop a more effective crowd counting method by utilizing temporal information in video sequences.
• To improve performance by imposing stronger constraints based on people conservation.

Main Methods:

• Estimating people flow between consecutive frames instead of directly regressing densities.
• Imposing strong people conservation constraints.
• Exploiting the correlation between people flow and optical flow.
• Utilizing people conservation in spatial and temporal domains for active learning.

Main Results:

• Significantly boosted performance in crowd counting without architectural complexity.
• Achieved further improvements by leveraging the correlation with optical flow.
• Enabled training of deep crowd counting models with significantly fewer annotations in an active learning setting.

Conclusions:

• Estimating people flow and enforcing conservation offers a more robust approach to crowd counting.
• This method reduces annotation costs while maintaining high performance.
• The approach enhances the efficiency and effectiveness of crowd counting systems.