Smart Camera Aware Crowd Counting via Multiple Task Fractional Stride Deep Learning.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Estimating crowd counts in dense, occluded scenes is difficult.
• Traditional methods use CNNs for density map regression, but can lose detail.

Purpose of the Study:

• To develop a deep learning model for accurate crowd counting.
• To simultaneously predict density maps and classify crowd density levels.
• To prevent dangerous stampedes using smart camera technology.

Main Methods:

• Proposed a Convolutional Atrous Fractional Network (CAFN) for larger receptive fields and reduced detail loss.
• Developed a Multiple Tasks CAFN (MTCAFN) for simultaneous density map regression and density level classification.
• Utilized dilated kernels and fractional strides in the convolutional neural network architecture.

Main Results:

• The MTCAFN model demonstrated effective performance on four benchmark datasets.
• Achieved Mean Absolute Errors (MAE) of 88.1 (Shanghai Tech A), 18.8 (Shanghai Tech B), 8.2 (WorldExpo'10), and 303.2 (NS UCF_CC_50).

Conclusions:

• The proposed MTCAFN method offers an effective approach for crowd counting and density level classification.
• The dual-task approach helps relax network parameters, contributing to safety applications like preventing stampedes.