Visual Analysis of Extremely Dense Crowded Scenes

Explore advanced computer vision techniques for analyzing extremely dense crowded scenes in this doctoral dissertation defense. Delve into novel approaches for counting, detecting, and tracking individuals in images and videos containing hundreds or thousands of people. Learn about methods combining low confidence head detection, texture repetition analysis, and frequency-domain processing. Discover how Markov Random Fields are employed to account for count disparities across scales and neighborhoods. Examine the use of binary integer least squares with Special Ordered Set Type 1 constraints for hypothesis selection. Investigate context-aware human detection in low to medium density crowds using locally-consistent scale priors. Study tracking techniques for dense crowds, including the identification of prominent individuals and the application of Neighborhood Motion Concurrence to model crowd behavior. Gain insights into cutting-edge research addressing challenges in computer vision and crowd analysis through detailed explanations, methodologies, and experimental results presented in this comprehensive oral examination.

Motivation
Presentation Layout
Key Ideas
Problems
Related Work: Counting by Detection
Related Work: Counting by Regression
Spatial Poisson Counting Process
Patches: Head Detections
Patches: Fourier Analysis
Patches: Interest Points
Patches: Fusion
Images: Multi-scale MRF
Results: Quantitative
Results: Per Patch Analysis
Results: Performance Analysis
Results: Analysis of 10th Group
Localization
Schematic Outline
Search results: Uniform Grid
Finding Representative Templates
Hypotheses Selection
Optimization
Bint Quadratic Programming
Background: Deformable Parts Model
Framework
Scale and Confidence Priors
Intermediate Results
Combination-of-Parts (COP) Detection
Global Occlusion Reasoning
Dataset: UCF-HDDC
Results: Qualitative
Results: Step-wise Improvement
Results: Density based Analysis
Results: Comparison
Results: Failure Cases
Chapter Summary
Queen Detection
Detection of Prominent Individuals
Modeling Crowd Behavior
Neighborhood Motion Concurrence
Tracking: Hierarchical Update
Experiments: Sequences
Quantitative Comparison
Component Contribution
Dissertation Conclusion
Future Work