Embed Size (px)
Citation preview
1
Associating People Dropping
off and Picking up Objects
Dima Damen, David HoggComputer Vision Group – University of Leeds
BMVC07
2/25BMVC 07
The Task
3/25BMVC 07
Tracking people and detecting objects
The video is a sequence of periods of activity
4/25BMVC 07
Tracking people and detecting objects
The video is a sequence of periods of activity
5/25BMVC 07
Associating drops with picks
6/25BMVC 07
Discriminating drops from picks - people
7/25BMVC 07
Discriminating drops from picks - people
8/25BMVC 07
Discriminating drops from picks - objects
Masked edges
‘before’ reference image
‘after’ reference image
9/25BMVC 07
Possible assignments for a period of activity
Assignments set 1:
drop(Person1, Bike1)
none(Person2)
drop(Person3, Bike1)
drop(Person4, Bike1)
none(Person5)
drop(Person6, Bike2)
Person
1
2
3
4
5
6
Bike
1
2
10/25BMVC 07
Possible assignments for a period of activity
Assignments set 1:
drop(Person1, Bike1)
none(Person2)
drop(Person3, Bike1)
drop(Person4, Bike1)
none(Person5)
drop(Person6, Bike2)
Person
1
2
3
4
5
6
Bike
1
2
11/25BMVC 07
Tree of hypotheses: sequences of assignments
Period of activity
1
2
12/25BMVC 07
Tree of hypotheses: sequences of assignments
Period of activity
1
2
13/25BMVC 07
Tree of hypotheses: sequences of assignments
Period of activity
1
2
14/25BMVC 07
Tree of hypotheses: sequences of assignments
Period of activity
1
2
3
15/25BMVC 07
Tree of hypotheses: sequences of assignments
Period of activity
1
2
3
K-best – k-min-cost
16/25BMVC 07
Tree of hypotheses: sequences of assignments
Period of activity
1
2
3
K-best – k-min-cost
17/25BMVC 07
Tree of hypotheses: sequences of assignments
Period of activity
1
2
3
K-best – k-min-cost
18/25BMVC 07
Tree of hypotheses: sequences of assignments
19/25BMVC 07
Tree of hypotheses: sequences of assignments
20/25BMVC 07
Constrained optimisation
21/25BMVC 07
Post-segmentation
22/25BMVC 07
Experiments & Results
� 3 experiments
� 1 hour (45 events)
� 50 minutes (22 events)
� Full day (9 hours and 30mins) (40 events)
% of correct connections
83.59
70.37
75.86
Unconstrained
96.093
92.592
93.101
ConstrainedExp #
23/25BMVC 07
Application to bicycle theft detection
� 8×8×8 scale-normalized equal-bin-size colour
histogram
� Scale-by-max
� Median histogram
Predicted
18317Non-Thief
310Thief
Non-ThiefThiefActual
24/25BMVC 07
Summary
� Deal with ambiguity in the visual data through the use of global constraints on what is possible.
� Comparison with unconstrained and partially-constrained solutions (in the paper).
� Ambiguities in the observations are expressed as multiple hypotheses.
� Hypotheses can then be verified or invalidated by future observations.
25/25BMVC 07
Thank you for listening
Contact Info:Dima DamenComputer Vision Group – University of [email protected]
