Recognizing Behaviors and Actions in Video

1
Recognizing Behaviors and Actions in Video

• Professor Robert Pless
• Department of Computer Science and Engineering

2
Continued work in real-world surveillance

• Today
• Anomaly detection -- Using motion patterns as
  background models.
• Recognizing features of background models.
• Interpretation of video deformations.

3
Anomaly Detection
Over what time scale are variations
relevant? This is a short time scale
4
How short of a time scale?
5
Anomaly detection
6
Longer time scales
Vandal Real time system on standard laptop
Kingshighway and Lindell intersection, St. Louis,
MO (Chase Park Plaza Hotel)
7
Background Model
Generic algorithm Build Model of Typical
Background Motion. Mark areas of video that
dont fit that model
8
Unusual Traffic Motion
Independent Motion Detection with temporal
integration
9
More than a background model.
Is this more than a background model? Statistical
ly consistent motion patterns define areas of
consistent motion on image. This is, in a sense,
a map defined by functional properties.
10
(No Transcript)
11
Automatic map generation
Very preliminary so far. Next step Map
generated from motion pattern interpretation may
bootstrap appearance based models of roads.
12
Video Analysis Reprise
Image similarities define similarity structures
in the video.
13
More analytic results application to MRI images.
Cardiac MRI, courtesy of Nikos Tsekos, Department
of Radiology, Washington University Medical
School.
14
Better Distance Function Gives Better Results.
Motion axis
Motion axis
Contrast change
15
Automatic Registration of complex deformation.
16
Actively Seeking

• Test Data for all of the above algorithms
• Anomaly detection, motion based feature
  recognition, and deformation analysis.
• MRI demo is live online at
• www.cse.wustl.edu/pless/wumapDemo