Locating and Counting Heads in Crowds With a Depth Prior

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Crowd counting and localization are challenging computer vision tasks, especially in dense or low-visibility scenarios.
• Existing detection-based methods struggle with accurately identifying small or occluded heads.
• RGB-D data offers depth information that can significantly improve crowd analysis.

Purpose of the Study:

• To develop a robust detection-based crowd counting method using RGB-D data.
• To enhance the accuracy of head detection and localization, particularly for dense and small targets.
• To address limitations in existing RGB-D datasets for crowd counting research.

Main Methods:

• Designed a dual-path guided detection network (DPDNet) leveraging RGB-D data.
• Introduced a density map guided detection module with a depth-adaptive kernel for improved head classification and density map regression.
• Proposed a density map guided Non-Maximum Suppression (NMS) strategy to prevent filtering of dense heads.
• Developed a depth-guided detection module with dynamic dilated convolution and depth-aware anchors for detecting small heads.
• Collected two large-scale RGB-D crowd counting datasets (synthetic and real-world) and proposed a meta-learning-based depth completion method.

Main Results:

• DPDNet achieved state-of-the-art performance on RGB-D crowd counting and localization tasks across multiple datasets.
• The proposed density map guided modules and depth-aware components significantly improved detection accuracy for dense and small heads.
• The method demonstrated effective performance even when extended to RGB-only crowd counting scenarios.

Conclusions:

• The DPDNet effectively addresses the challenges of crowd counting and localization using RGB-D data.
• The proposed techniques for handling dense/small heads and depth information are crucial for robust crowd estimation.
• The developed datasets and depth completion method facilitate future research in RGB-D crowd counting.