A Dilated Convolutional Neural Network for Cross-Layers of Contextual Information for Congested Crowd Counting

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Crowd counting is crucial for many applications, with convolutional neural networks (CNNs) showing promise.
• Existing CNN-based methods often overlook the significance of shallow features, focusing primarily on deep feature maps.
• This limitation hinders optimal performance in complex crowd density estimation tasks.

Purpose of the Study:

• To propose a novel network, the cross-level contextual information extraction network (CL-DCNN), for crowd counting.
• To effectively integrate both deep and shallow features by extracting cross-level contextual information.
• To enhance the accuracy and robustness of crowd counting models.

Main Methods:

• Development of a dilated convolutional-neural-network-based approach (CL-DCNN).
• Introduction of a dilated contextual module (DCM) to connect different feature maps.
• Integration of cross-level connections to leverage multi-level feature information for crowd scene analysis.

Main Results:

• The proposed CL-DCNN effectively integrates contextual information while preserving local details of crowd scenes.
• Experiments on five public datasets (ShanghaiTech A/B, Mall, UCF_CC_50, UCF_QNRF) demonstrate superior performance.
• Achieved Mean Absolute Errors (MAE) of 52.6, 8.1, 1.55, 181.8, and 96.4 on the respective datasets.

Conclusions:

• The CL-DCNN approach significantly outperforms state-of-the-art methods in crowd counting.
• The integration of shallow and deep features through cross-level contextual information is key to improved performance.
• The proposed method offers a more effective solution for accurate crowd density estimation.