An Adaptive Multi-Scale Network Based on Depth Information for Crowd Counting.

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Crowd counting is crucial for public safety and intelligent transportation but faces challenges like occlusion, inconsistent target sizes, and annotation errors.
• Existing density plot regression methods struggle with scale variations and distinguishing near/far targets, leading to poor performance in sparse areas.

Purpose of the Study:

• To propose an adaptive multi-scale network for accurate and efficient crowd counting.
• To address limitations in existing methods regarding scale adaptability and feature distinction between near and far targets.

Main Methods:

• Developed an adaptive multi-scale far and near distance network using a convolutional neural network (CNN) framework.
• Employed stacked convolution layers for deeper networks, allocated different receptive fields based on target distance, and fused nearby features.
• Integrated depth information and pixel-level adaptive modeling by dividing images into patches.
• Utilized density normalized average precision (nAP) for spatial positioning accuracy analysis.

Main Results:

• The proposed network achieves a good balance between accuracy, inference speed, and performance.
• Demonstrated significant performance improvements over State-Of-The-Art (SOTA) methods on challenging benchmarks (Shanghai Tech A/B, UCF_CC_50, UCF-QNRF).
• Effectively handled complex backgrounds and diverse crowd distributions, including severe occlusion and scale variations.

Conclusions:

• The adaptive multi-scale network provides a robust solution for crowd counting in various scenarios.
• The method enhances feature extraction and adaptively models populations, overcoming limitations of traditional approaches.
• Validated effectiveness on multiple datasets, showing superior performance and robustness against complex environmental factors.