Redesigned Skip-Network for Crowd Counting with Dilated Convolution and Backward Connection.

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Crowd counting faces challenges due to variations in object scale and density.
• Existing methods often prioritize high-level features, neglecting crucial low-level details.

Purpose of the Study:

• To propose an enhanced crowd counting estimation network.
• To improve the emphasis on low-level features within a hierarchical network structure.

Main Methods:

• Developed an estimation network integrating high-level features into shallow layers.
• Utilized dilated convolution to preserve semantic information without resizing feature maps.
• Employed two identical networks for feature extraction and final result estimation.

Main Results:

• The proposed network demonstrated superior performance in high crowd density conditions.
• Achieved significant reduction in over-counting errors across tested datasets.
• Evaluated using mean absolute error and root mean squared error for accuracy and robustness.

Conclusions:

• The enhanced network effectively addresses limitations of existing crowd counting methods.
• The approach offers improved accuracy and robustness, particularly in dense crowd scenarios.
• This work contributes to more reliable crowd analysis in computer vision applications.