Human Identification From a Distance D. Adjeroh, B. Cukic, M. Gautam, L. Hornak, A. Ross Lane Department of CSEE West Virginia University NC-BSI, December 2008
NC – BSI Problem Statement Surveillance datasets acquired at border zones offer an opportunity to recognize individuals from a distance rather than requiring close visual inspection. The project will develop methods for human identification from surveillance videos. Methodology Develop a hierarchical approach to human recognition from a distance. Define event clustering in joint biometric – surveillance space. Search methods: from events to biometric profiles and vice versa.
NC – BSI Image Quality at a Distance High Sensitivity to Motion blur: because of long focal distance Out-of-focus Blur: because of small DOF Distortion due to lens Low pixel count: (sensor resolution is limited) Magnification blur (due to high magnification) (66×, 50m) (109×, 100m)(153×, 150m)(284×, 300m) Note: (magnification, distance) approximately the same resolution: 60 pixels between the eyes.
NC – BSI Surveillance Applications Outdoor Location 100x – 100m75x – 50m 200x – 200m300x – 300m Range finder Telescope
NC – BSI Effect of Frame Resolution Rank 1 CMC curves 60 pixels 35 pixels 85 pixels 10x, 52f10x, 31f15x, 31f
NC – BSI Effect of Illumination CMCs Rank 1 100% roof light50% roof lightNo roof light Degradations in high magnification images: Sensor noise Magnification blur Motion blur Out of focus blur Zoom blur Atmospheric blur Illumination Contrast Resolution Probes: 20x magnification 52 feet, 50 pixels inter-eye distance
NC – BSI Soft biometric traits Jain et al, “Utilizing soft biometric traits for person authentication”, Proc. International Conference on Biometric Authentication (ICBA), Hong Kong, July 2004
NC – BSI Combining Face & Soft Biometrics
NC – BSI Human Metrology 2D Model –Available from video –Possible multiple views in surveillance MAT Representation –Medial Axis Transform –Less detailed, but may be adequate for required representation Decorated MAT Representation (for 2D) MAT Representation (1D) Multiresolution MATs 2D measurements superimposed on 3D images (3D images from Allen et al, 2004)
NC – BSI Extending the Application Envelope: Virtual Identities in Space/Time Correlate Two Surveillance Videos Between Aldgate East and Liverpool Street tube stations Between Russell Square and King's Cross tube stations At Edgware Road tube station On bus at Tavistock Square
NC – BSI Extending the Application Envelope (2)
NC – BSI Biometric Surveillance Space
NC – BSI Decomposing a Video Stream
NC – BSI Retrieval/Analysis Paradigms
NC – BSI Leverage The Center for Identification Technology Research (NSF I/UCRC). Biometrics: Performance, Security and Social Impact, (NSF and DHS – Human Factors) Biometric recognition from video streams, data collection. Night time biometrics (ONR). Video/image compression.
NC – BSI Deliverables Years 2-6: Architecture of the joint identity-surveillance space, Efficient segmentation and labeling algorithms, Fusion algorithms for identification from surveillance video, Storage and retrieval architecture. System evaluation.