Artificial Vision LOGICAL ARCHITECTURE Dr. Christian Micheloni Department of Computer Science University of Udine, ITALY Artificial Vision State of the art • First Generation of Surveillance Systems with analog data Transmisison (1960-1980). Control center Multiplexer Analog transceiver Sensor level Local processing level Klagenfurt 6-11 April Network level (coax 75 ohm cable network) 2011 PAGE 2 Analog storage devices Prof. Micheloni Christian Università Degli Studi di Udine Artificial Vision Second Systems Generation (1990-2000) State of the art (2) IEEE 802.11 Cable modem HDSL modem Sensor level Signal Processing Network Local signal processing (channel HUB WAN o LAN coding) Network level Klagenfurt 6-11 April PC’s Cluster (attention driving) 2011 PAGE 3 Operator level Prof. Micheloni Christian Università Degli Studi di Udine Artificial Vision Third Systems Generation (2000- ???) State of the art (3) UMTS Cable modem ADSL modem Sensor Network HUB level Video processing and recognition at the sensor level Klagenfurt 6-11 April Local signal processing (channel WAN o LAN Network level coding) 2011 PAGE 4 Decision planning Operator level Prof. Micheloni Christian Università Degli Studi di Udine Artificial Vision Logical architecture Image Acquisition Frame Grabber HW Frame Grabber SW Frame Corrente Creation and Updating of Background Change Detectioni Motion Analysis Focus of Attention Tracking Feature extraction Localization GRAPHIC INTERFACE Motion Detection Event Analysis Identification Database Classification ALARM Klagenfurt 6-11 April 2011 PAGE 5 Prof. Micheloni Christian Università Degli Studi di Udine Artificial Vision Background Updating Change Detection Bck (t ) Current Frame I (t ) Change Detection Delay Delay I (t 1) Klagenfurt 6-11 April I t 1 Bck t 1 I t 1 Bck (t 1) Background Updating 2011 Prof. Micheloni Christian Università Degli Studi di Udine Artificial Vision Mobile object detection - Detection - Localization - Tracking Klagenfurt 6-11 April 2011 PAGE 7 Prof. Micheloni Christian Università Degli Studi di Udine Artificial Vision Object detection Change Detection Klagenfurt 6-11 April 2011 PAGE 8 Prof. Micheloni Christian Università Degli Studi di Udine Artificial Vision Object detection (2) Change Detection Klagenfurt 6-11 April 2011 PAGE 9 Prof. Micheloni Christian Università Degli Studi di Udine Artificial Vision Object localization Klagenfurt 6-11 April 2011 PAGE 10 Prof. Micheloni Christian Università Degli Studi di Udine Artificial Vision Object tracking • It consists in the temporal association of the objects A' A B B' C C' tk D' D D(A,A’) = 2.03 D(A,B’) = 0.39 D(A,B’) = 0.39 D(A,B’) = 0.39 Klagenfurt 6-11 April A -> A’ D(B,A’) = 0.23 B -> B’ D(B,B’) = 2.0 2011 PAGE 11 D(C,A’) = 0.03 D(C,C’) = 2.39 tk+N C -> C’ D(D,A’) = 0.0023 D -> D’ D(D,D’) = 4.0 Prof. Micheloni Christian Università Degli Studi di Udine Artificial Vision Object tracking (2) Klagenfurt 6-11 April 2011 PAGE 12 Prof. Micheloni Christian Università Degli Studi di Udine Artificial Vision Object classification PERSON VEHICLE GROUP OF PEOPLE Neural network Klagenfurt 6-11 April 2011 PAGE 13 Prof. Micheloni Christian Università Degli Studi di Udine Artificial Vision Object classification (2) Klagenfurt 6-11 April 2011 PAGE 14 Prof. Micheloni Christian Università Degli Studi di Udine Artificial Vision Object tracking and classification Klagenfurt 6-11 April 2011 PAGE 15 Prof. Micheloni Christian Università Degli Studi di Udine Artificial Vision Event analysis and classification Event 1 Video Flow Features extraction for event analysis Algorithm for event classification Event 2 . . . Event N Feture Modelling pattern Clustering and Timing • The selected features for event analysis are : • Tracks • Starting and ending time • Class of the object Klagenfurt 6-11 April 2011 PAGE 16 Prof. Micheloni Christian Università Degli Studi di Udine Artificial Vision Explicit Event Definition An anomalous event can be identified on the bases of its explicit definition. Person walks Car 2 exits Car 1 enters Klagenfurt 6-11 April 2011 PAGE 17 Prof. Micheloni Christian Università Degli Studi di Udine Artificial Vision Explicit Event Definition (2) Event 1 Event 2 Event 3 • The sequential recognition of the events 1,2 and 3 could be interpreted as the accomplishment of an anomalous activity. • Not always the anomalous events are a simple sequence of simple events Klagenfurt 6-11 April 2011 PAGE 18 Prof. Micheloni Christian Università Degli Studi di Udine Artificial Vision Probabilistic Events Analysis • The explicit definition of anomalous events is sometimes to rigorous and it does not fit well with the dynamics in the scene. • The explicit definition could be difficult to perform (either manually or with automatic learning mechanisms) Alternative: • To adopt a probabilistic approach: an unlikely event with respect to the detected event is also an anomalous event. Klagenfurt 6-11 April 2011 PAGE 19 Prof. Micheloni Christian Università Degli Studi di Udine Artificial Vision STATE of the ART • G.L. Foresti, P. Mahonen and C.S. Regazzoni, Multimedia VideoBased Surveillance Systems: from User Requirements to Research Solutions, Kluwer Academic Publishers, 2000. • L. Marcenaro, F. Oberti, G.L. Foresti and C.S. Regazzoni, “Distributed Architectures and Logical Task Decomposition in Multimedia Surveillance Systems”, Proceedings of the IEEE, Vol. 89, no. 10, October 2001, pp. 1419-1440. • G.L. Foresti, C. Micheloni, L. Snidaro, P. Remagnino and T. Ellis, “Active Video-based Surveillance Systems: the low-level image and video processing techniques needed for implementation”, IEEE Signal Processing Magazine, Vol. 22, No. 2, March 2005, pp. 25-37. Klagenfurt 6-11 April 2011 PAGE 20 Prof. Micheloni Christian Università Degli Studi di Udine