Detecting Humans in Dense Crowds Using Locally-Consistent Scale Prior and Global Occlusion Reasoning

Area of Science:

• Computer Vision
• Artificial Intelligence
• Pattern Recognition

Background:

• Human detection in dense crowds is crucial for various visual analysis tasks.
• Challenges include scale variation, occlusion, and perspective distortion.
• Crowded scenes offer contextual cues that can aid detection.

Purpose of the Study:

• To develop an effective method for human detection in dense crowds.
• To utilize contextual information, specifically locally-consistent scale priors.
• To improve occlusion reasoning and body part detection.

Main Methods:

• Inferred scale and confidence priors using Markov Random Field (MRF) from initial detections.
• Iteratively refined detection confidences and priors.
• Employed Binary Integer Programming (BIP) for occlusion reasoning.
• Utilized local neighbor-dependent constraints for detection and reasoning.
• Proposed a mechanism for detecting body part combinations without specific annotations.

Main Results:

• Demonstrated marked improvement over the underlying human detector on a challenging dense crowd dataset.
• Successfully addressed challenges of scale variation, occlusion, and perspective distortion.
• Enabled robust human detection and occlusion reasoning in dense crowd scenarios.

Conclusions:

• The proposed context-aware approach significantly enhances human detection in dense crowds.
• Leveraging locally-consistent scale priors and neighbor-dependent constraints is effective.
• The method shows promise for real-world applications requiring crowd analysis.