Motion Estimation and Object Tracking
Motion Estimation
Introduction
- Adding the time dimension to image formation
- Analyzing changing scenes via image sequences
- Changes in image sequences provide features for:
- Detecting moving objects
- Computing trajectories
- Performing motion analysis
- Recognizing objects based on behaviors
- Computing viewer motion
- Detecting and recognizing activities
Applications
- Motion-based recognition
- Automated surveillance
- Video indexing
- Human-computer interaction
- Traffic monitoring
- Vehicle navigation
Scenarios
- Still camera
- Constant background with single/multiple moving objects
- Moving camera
- Relatively constant scene with coherent motion or moving objects
Change Detection
- Detect moving objects against a constant background
- Steps:
- Derive background image
- Subtract background from each frame
- Threshold and enhance difference image
- Detect bounding boxes
Sparse Motion Estimation
- Compute sparse motion field by identifying corresponding points in two images
- Steps:
- Detect interesting points (using edge/corner detectors, SIFT, etc.)
- Search for corresponding points in the next frame
Dense Motion Estimation
- Optical Flow
- Assumptions:
- Object reflectivity and illumination don't change
- Distance to camera doesn't vary significantly
- Small neighborhoods shift position between frames
Optical Flow Equation
- Derived from Taylor series expansion
- Constraint:
- Requires additional constraints for unique solution
Lucas-Kanade Approach
- Assumes constant flow in local neighborhood
- Solves system of equations using least squares
Object Tracking
Introduction
- Generating inference about object motion from image sequences
Applications
- Motion capture
- Recognition from motion
- Surveillance
- Targeting
Challenges
- Loss of information in 2D projection
- Image noise
- Complex object motion
- Non-rigid objects
- Occlusions
- Complex shapes
- Illumination changes
- Real-time requirements
Bayesian Inference for Tracking
- Three main steps:
- Prediction
- Association
- Correction
Independence Assumptions
- Current state depends only on immediate past
- Measurements depend only on current state
Tracking Process
- Prediction:
- Correction:
Kalman Filtering
- Assumes linear models and Gaussian noise
- Prediction and correction steps with matrix operations
Particle Filtering
- For non-linear/non-Gaussian cases
- Represents state density with weighted particles
- Propagates particles using dynamics model
- Updates weights using measurement model
Applications
- Tracking active contours
- Object location tracking in clutter