MH-MetroNet-A Multi-Head CNN for Passenger-Crowd Attendance Estimation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Transportation Systems

Background:

• Accurate passenger attendance estimation in metro cars is vital for safe crowd coordination and management in stations.
• Existing crowd counting methods often struggle with dense and occluded scenes typical in public transport.

Purpose of the Study:

• To propose a novel multi-head Convolutional Neural Network (CNN) architecture, MH-MetroNet, for accurate passenger attendance estimation in metro cars.
• To evaluate the performance of MH-MetroNet against state-of-the-art crowd counting architectures on diverse datasets.

Main Methods:

• Development of a multi-head CNN architecture featuring a convolutional backbone for feature extraction and multi-head layers for density map estimation.
• Training and testing the MH-MetroNet on publicly available crowd counting datasets (ShanghaiTech A/B, UCF_CC_50) and a custom dataset from Italian subway cars.
• Comparison of MH-MetroNet's performance using Mean Absolute Error (MAE) and Mean Square Error (MSE) against established crowd counting models.

Main Results:

• MH-MetroNet demonstrated superior performance in predicting passenger counts compared to existing state-of-the-art crowd counting architectures.
• The proposed architecture achieved lower Mean Absolute Error (MAE) and Mean Square Error (MSE) on both benchmark and real-world subway datasets.
• Effective estimation of crowd density maps was achieved, leading to accurate people number prediction.

Conclusions:

• The MH-MetroNet architecture offers a robust and accurate solution for estimating passenger attendance in metro environments.
• This technology can significantly enhance the safety and efficiency of crowd management in metro systems.
• The study validates the effectiveness of the proposed CNN approach for real-world crowd counting applications in transportation.