Video Event Detection
Techniques
Filip Nálepa, Michal Batko
Outline
● Motivation
● IBM Smart Surveillance System
● Background subtraction
● Appearance-based tracking
● Our approach
● Optical flow
● Comparison
Motivation
● Automatic surveillance video analysis
● Real-time alarms for suspicious behaviour
● Situation awarness
IBM Smart Surveillance
System
● Real-time alerts
● Indexed event search
● Approach
● Object detection – background subtraction
● Object tracking – appearance-based
● Object classification
● Event detection
Background Subtraction
● Moving objects detection
● Separation of foreground and background
● Background model
● Adaptive
● Multimodal
Appearance-based Tracking
● Associating objects in successive frames
● Based on distance and areas
● Appearance model for each object –
resolving splits and merges
Object Classification
● Person or vehicle
● Object size
● Velocity direction
● Object location
Our approach
● Real-time event detection
● Approach
● Object detection – face detector
● Object tracking – optical flow
● Event detection
Optical Flow
● Motion vectors
● Assumptions
● Brightness constancy
● Small movements
● More assumptions needed
● Dense and sparse techniques
Good Features to Track
● Corners
● Harris
● Shi and Tomasi
Lucas-Kanade
● Sparse optical flow technique
● Windows matching
● Assumptions
● Brightness constancy
● Small movements
● Spatial coherence
● Works for corners
Spatial coherence
Pyramidal Lucas-Kanade
● Large movements tracking
● Resolution levels
● Small and coherent motion
violations mitigated
Object Detection Comparison
● Face detector
● Face detection only
● No adaptation
problems
● 1 frame needed
● No further
classification
● Background subtraction
● Any moving object
detected
● Adaptation problems
● More frames needed
● Object classification
needed
Object Tracking Comparison
● Optical flow
● Motion vectors
computation
● Tracked points
● Points loss → object
loss
● Works on camera
motion
● Background subtraction
and appearance-based
tracking
● Objects associating
● Entire image
● Occlusion handling
● Just static camera
Object Classification
Comparison
● Face detector
● Face detection only
● No configuration
● SSS approach
● Multiple object
types
● Configuration
needed
Possible Combination
● Object detection – background subtraction
● Tracking – optical flow (+ appearance-
based)
● Object classification – face detector and
other object features
Summary
● IBM Smart Surveillance System
● Background subtraction
● Appearance-based tracking
● Our approach
● Face detector
● Optical flow
Thank you for your attention.