Detection And Tracking Of multiple Targets In Crowded Scenes

Open Access
- Author:
- Degol, Joseph Michael
- Area of Honors:
- Computer Engineering
- Degree:
- Bachelor of Science
- Document Type:
- Thesis
- Thesis Supervisors:
- Dr. Robert Collins, Thesis Supervisor
Lee David Coraor, Thesis Honors Advisor - Keywords:
- Mean Shift Belief Propagation
Tracking
Detection
Feature Points
Computer Vision
Computer Science
Artificial Intelligence - Abstract:
- In this work, we propose a method for detecting and tracking groups of indi- viduals in crowded scenes. Our method begins with a detection phase where we use KLT to extract feature points from the scene. Feature Points that are stationary or have infeasible trajectories are then pruned leaving only points representing moving people. These points are then connected to form a graph structure using Delaunay Triangulation. Edges of the graph are as- signed weights based on the proximity and coherency of motion of connected feature points. Edge weights above a given threshold are then pruned, result- ing in blobs of connected feature points. These blobs represent our groups of individuals that we want to track; this concludes the process of detecting groups of individuals. After the detection phase is complete, we use mean shift belief propagation (MSBP) to track the feature points. This method allows connected feature points to in uence the likelihood estimation for each point, resulting in an estimation that is based on the movement of the en- tire group of individuals. In the past, MSBP has been used for tracking of rigid lattice graphs, but not for general graph structures. Thus, much of the novelty of this approach comes from the application of MSBP to track connected feature points which are represented by general graph structures. Additionally, this paper demonstrates a vision system that combines estab- lished techniques in a novel way for both detection and tracking.