Explore a doctoral dissertation defense on advanced multi-target tracking techniques for pedestrians. Delve into novel data association methods using Generalized Maximum Clique Problem (GMCP) and Generalized Maximum Multi Clique Problem (GMMCP) formulations. Learn about global optimization approaches incorporating motion and appearance, and discover solutions for tracking in extremely crowded scenes using Binary Quadratic Programming. Examine the limitations of current methods, proposed improvements, and experimental results across various datasets. Gain insights into cutting-edge algorithms for efficient tracking of hundreds of targets in complex scenarios like marathons and political rallies.
Overview
Syllabus
Intro
Challenges
Multi-target Tracking: Applications
Outline
Data Association
GMCP Tracker: Pipeline
How to solve GMCP?
Process of Finding Tracklets in one Segment
Parking Lot Results
Evaluation Metrics
Limitations
What are the main differences?
Framework
Mid-level Tracklet Generation
Optimization
Aggregated Dummy Nodes (ADN)
Run-time Comparison
Qualitative Results
Parking Lot 2
Occlusion Handling
Quantitative Comparison
Crowd Tracking
Spatial Proximity Constraint
Neighborhood Motion Effect
Grouping
Formulation
Appearance
Quadratic Constraints
Frank Wolfe Algorithm
Frank Wolfe with SWAP steps
Experiments . 9 high-density sequences
Quantitative Results
Contribution of each term
Summary
Future Work
Taught by
UCF CRCV
