Vision Surveillance Paul Scovanner
Surveillance Main tasks Locating people and objects in a scene Background Subtraction Object Detection Track objects as they move Associate objects across frames Beyond Tracking
Background Subtraction Remove the background leaving areas where movement occurs Frame Differencing: |framet – framet-1| > Threshold
Background Subtraction Frame Differencing Fast Simple Error prone (Illumination changes, Edges on large objects, Amplifies sensor noise) Background Modeling |framet – Background| > Threshold Model the colors of each pixel as a Gaussian (mean and standard deviation) Still cant detect stationary objects
Background Subtraction Mixture of Gaussians
Object Detection aka “is that a car or a person?” Aspect ratio Object Detectors
Tracking We can detect moving objects (If background subtraction works) We can identify pedestrians and cars (If object detection works) What’s left?
Tracking Associate the detections in one frame with the next. Visual similarity Spatial location
Tracking
Multi-view Tracking If 1 camera is good… 3 Must be better
Multi-view Tracking
Multi-view Tracking
Tracking From The Air
Tracking From The Air
Tracking From The Air
Tracking Prediction Pedestrian Modeling Predict movements of pedestrians
Anomaly Detection Detect emergency events
Anomaly Detection
More Than Just Tracking Crowd instability