Suspicious Behavior in Outdoor Video Analysis Challenges

1
Suspicious Behavior in Outdoor Video Analysis -
Challenges Complexities

• Air Force Institute of Technology/ROME Air Force
  Research Lab
• Unclassified IED test sequences showing dropped
  package from vehicle (DPV)
• Combination of motion analysis and change
  detection
• Homogenous regions and aperture problem and for
  optic flow approaches
• Learning appropriate background for change (ghost
  objects appear due to slow or fast learning)
• Global camera motion/jitter
• Occlusion and Camouflage
• Environmental problems
• Dust and smoke
• Wind local object motion (swaying of branches,
  shadows)
• Precipitation rain, slow etc.
• Clutter (background model)
• Illumination problems
• Shadows (static and moving cast shadow) - missed
  objects or false detections
• Glare false detections, object shape and
  trajectory distortions
• Sudden illumination changes (cloud movements)
  false detections
• Low contrast or color saturation

2
Dropped package from vehicle (DPV) sequences -
Logitech Orbit 20ft Vertical Run 2
Frame 150 Event of interest marked
3
Detecting Occluded EventSequence - Logitech
Orbit 20ft Vertical Run 3
Frame 121 Event of interest marked
4
Glare and ShadowsSequence - Logitech Orbit 10ft
Vertical Run 2
Frame 43 False detection due to glare
Frame 141 Event of interest marked
5
Dust and ShadowsSequence - Logitech Orbit 10ft
Vertical Run 3
Frame 159 Event of interest marked
Frames 72170 False detection due to dust
6
Dust and ShadowsSequence - Logitech QuickCamPro
5000 10ft Vertical Run 3
Frame 91 False detections due to dust
Frame 150 Event of interest missed due to
shadow and insufficient contrast
7
Sequence - Logitech QuickCamPro 5000 20ft
Vertical Run 2
No event of interest
8
Sequence - Logitech QuickCamPro 5000 20ft
Vertical Run 3
Frame 104 Event of interest marked
9
Effect of Learning Rate in Background Modeling
Sequence - Logitech Orbit 10ft Vertical Run 2
Frame 86 Correct Detection when in motion
Frame 210 Ghost object left behind (due to slow
background learning using Mixture of Gaussians)
when the car starts to move again
Frame 134 Object that stops for a while blends
into the background
10
Problems with Flow-based Approaches Sequence -
Logitech QuickCamPro 5000 20ft Vertical Run 1
Frame 24 Aperture problem, motion of
homogeneous regions is not detected
Frame 83 Larger temporal window results in
false detections and larger object boundaries
Frame 35 Non-moving objects not detected
11
Suspicious Behavior in Outdoor Video Analysis -
Challenges Complexities

• Combination of motion analysis and change
  detection
• Homogenous regions and aperture problem and for
  optic flow approaches
• Learning appropriate background for change (ghost
  objects appear due to slow or fast learning)
• Global camera motion/jitter
• Occlusion and Camouflage
• Environmental problems
• Dust and smoke
• Wind local object motion (swaying of branches,
  shadows)
• Precipitation rain, slow etc.
• Clutter (background model)
• Illumination problems
• Shadows (static and moving cast shadow) - missed
  objects or false detections
• Glare false detections, object shape and
  trajectory distortions
• Sudden illumination changes (cloud movements)
  false detections
• Low contrast or color saturation

12
Moving Object Detection Approaches

• Optical Flow Analysis Characteristics of flow
  (velocity) vectors of moving objects over time
  are used to detect changed regions.
• Advantage can be used in the presence of camera
  motion.
• Disadvantage usually computationally expensive
  aperture problem.
• Change Detection
• Background subtraction Moving regions are
  detected through difference between the current
  frame and a reference background image.
• framei-Backgroundi gtTh
• Advantage provides the most complete feature
  data.
• Disadvantage sensitive to dynamic scene changes
  due to lighting and extraneous events and cannot
  handle global motion.
• Temporal differencing Similar to background
  subtraction but the estimated background is the
  previous frame.
• framei-framei-1 gtTh
• Advantage very adaptive to dynamic environments.
• Disadvantage has problems in extraction of all
  relevant feature pixels (aperture problem).