Respectful Cameras Jeremy Schiff EECS Department University of California, Berkeley Ken Goldberg, Marci Meingast, Deirdre Mulligan, Pam Samuelson IEOR, EECS, Law University of California, Berkeley http://www.cs.berkeley.edu/~jschiff/RespectfulCameras NSF Science and Technology Center, Team for Research in Ubiquitous Secure Technologies, NSF CCF-0424422, with additional support from Cisco, HP, IBM, Intel, Microsoft, Symmantec, Telecom Italia and United Technologies. Background New class of Robotic Cameras since 9/11/2001 $20,000 -> Under $1,000 Static -> Pan, tilt, zoom (21x) UK - 3 Million Outdoor Cameras Now Deploying in Large US Cities Zoom Example Invasiveness Objective Static Marker Detection Adaboost Training Phase Classifying Phase Input is data and label Data -> label Linear function of weak classifiers Example Construction Hat Color Features Input from images Each pixel red, green, blue (RGB) Values 0 to 255 Project into higher dimension Convert to 9 dimensions RGB HSV Stable over changing lighting LAB Good for detecting specularities 10 243 13 9 241 16 12 252 8 60 201 73 69 225 74 42 17 38 65 209 78 74 220 171 45 112 16 Classifiers Operates on each dimension Threshold value Above good and below bad Above bad and below good Example Connected Component Groups adjacent pixels Threshold Minimum Area Bounding Box Acceptable Ratio Between Dimensions Marker Tracking Particle Filtering Probabilistic Method for Tracking Motivates Probabilistic AdaBoost Particle filters Non-Parametric Sample Based Method (Particles) Particle Density ~ Likelihood Tracking requires three distributions Initialization Distribution Transition Model (Intruder Model) Observation Model Determines Observation Model 1-p p 0.1 0.1 0.1 0.2 0.0 0.8 0.6 0.4 0.2 0.7 0.9 0.4 0.3 0.2 0.1 0.2 0.9 0.9 0.9 0.8 1.0 0.8 0.6 0.6 0.8 0.7 0.9 0.6 0.7 0.8 0.9 0.8 0.79375 Transition Model State Position Bounding-box Width Bounding-box Height Orientation Speed Add Gaussian Noise to width, height, orientation and speed Euler Integration to determine new position Multiple Filters Single Filter Per Marker Define overlap Add Filter when overlap of Static Image Cluster and all filters is below threshold Delete Filter when prob. of best particle < 0.5 Delete Filter when 2 filters overlap > threshold Video – Nearby Hats Video – Nearby Hats Video – Lighting Video – Lighting Video – Crossing Video – Crossing Video – Shirt Video – Shirt Future Work Other Features Edge Detection Feature Structure Generalize to Other Domains Other Obstruction Mechanisms Encryption Full Body Multiple Cameras Thank You Jeremy Schiff: jschiff@cs.berkeley.edu URL: www.cs.berkeley.edu/~jschiff/RespectfulCameras