Transcript Document

GMCP-Tracker: Global Multi-object Tracking Using
Generalized Minimum Clique Graphs
UCF
Problem:

4.
Tracking by Detection
Key Contributions:

A new tool in Computer Vision:

Generalized Minimum Clique Problem (GMCP):
 Typically useful where there are multiple possibilities for some
subproblems, as well as a global criterion to satisfy.
 GMCP utilized for computing tracklets and trajectories.
A new temporally global approach to Data Association

A new tracklet-global motion cost model.

Shifting approximation from Time Domain to Object Domain:

Finding tracklets/trajectories in a temporally global way.

Finding the tracklet/trajectory of one object at a time (greedy).
𝑣21
𝑣12
𝐶2
𝑣22
𝑣14
𝐶5
𝑣24
𝑣23
𝑣32
𝑣31
𝐶4
Trajectories
𝑣13
𝐶3
𝑣15
𝑣25
𝑣34
𝑣35
𝐶6
ECCV ‘12
Found Tracklets in Different Segments:
Frame 3
𝑣33
𝑣𝟒6
𝑣16
𝑣26
𝑣36
5.
Frame 4
Frame 5
Frame 6
Frame 1
Frame 2
Frame 3
Input Graph G

•
Frame 2
𝑣11
𝐶1

({aroshan| adehghan | shah @cs.ucf.edu} , University of Central Florida)
Finding Tracklets Using GMCP:
Frame 1
Input Detections
1.
Amir Roshan Zamir, Afshin Dehghan, Mubarak Shah
Merging Tracklets into Trajectories:
Input Tracklets
•
Tracklet-global Motion Cost Model
6.
Experimental Results
Minimum Clique
Occlusion Handling using HN
Bipartite Matching vs. GMCP
Frame 5
Frame 6
Frame 1
Frame 2
Frame 3
Block Diagram:
3.
Generalized Minimum Clique Problem (GMCP):
Input to GMCP
Generalized Minimum Clique
Fra
me 2
Minimum Clique
2.
Frame 4
Fra
me 5
Frame 4
Frame 1

(nodes≡detections) (clusters≡frames) (GMCP solution=tracklet of one Object)
•
Occlusion Handling Using Hypothetical Nodes:
Project Page
Definition:
Frame 3
Project Page:
http://vision.eecs.ucf.edu/
projects/GMCP-Tracker/
YouTube

𝐺𝑠 = (𝑽𝑠 , 𝑬𝑠 , 𝑤𝑠 )
Frame 2
Frame 6
Occlusion Handling Using HN
𝐺 = (𝑽, 𝑬, 𝑤)
Frame 5
Frame 4
Frame 5
Frame 6