Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

Synesis Embedded Video Analytics


Published on

A set of new video analytics algorithms is described for automatic object detection and rule-based event recognition. The algorithms utilizes a 4D feature pyramid to model objects and the background in HD. A commercial version based TI's DaVinci DSP is embedded in intelligent IP-cameras and video encoders.

Published in: Technology
  • Be the first to comment

  • Be the first to like this

Synesis Embedded Video Analytics

  1. 1.<br />Embedded Video Analytics<br />DSP Algorithms forDetection, Tracking and Recognition<br />
  2. 2. HD Intelligent Network Video<br />Media and Internet<br />Face detection and recognition servers<br />Intelligent Video Surveillance<br />Intelligent cameras, encoders and DVRs<br />Digital TV<br />DVB receivers,STBs, PVRs,media centres<br />
  3. 3. What is the efficiency ofvideo surveillance?<br />Quality ofevent recognition<br />correct classification<br />response time<br />documentation<br />multiple locations<br />Operator comfort<br />Cost of ownership<br />
  4. 4. Video analytics and video analysis<br />?<br />
  5. 5. Functions of video analytics<br />Anti-tampering and operability monitoring<br />Operational alerts<br />Automatic priorities<br />Automatic PTZ-camera targeting<br />Event recording for instant forensic analysis<br />Optimal usage ofnetwork bandwidth and storage memory <br />
  6. 6. Solution: embedded video analytics<br />Edge device transmits video andmetadata (object and its behaviour description)<br />Zone 5intrusiondetected<br />VIDEO<br />EVENTDATABASE<br />EVENT RULES<br />METADATA<br />
  7. 7. Upon a suspicious event…<br />PTZ-targeting<br />System notificationover IP network to VMS<br />Sound and visual alarms, SMS etc<br />‘Dry contact’ signal<br />High quality recording to local or remote storage (NAS)<br />Analogue output to legacy systems (matrix or DVR)<br />
  8. 8. Embedded vs server analytics<br />BOTTLENECK<br />camera orencoder<br />video management system or DVR<br />compressedvideo & audio<br />Embedded(edge)analytics<br />codecs<br />video-analytics<br />video management system or DVR<br />камера или энкодер<br />Server(back-end)analytics<br />metadata<br />videoanalytics<br />video and audio<br />codecs<br />
  9. 9. Video signal sources<br />Network cameraAxis 211A<br />Analoguestandard definition cameras(PAL/NTSC)<br />Network cameras(standard and highdefinition)<br />Thermal cameras<br />Thermal cameraTitan-14<br />
  10. 10. Wide angle perimeter surveillance(multiple tripwire alert levels)<br />
  11. 11. Fence crossing detector<br />
  12. 12. Apartment housing event recording<br />
  13. 13. Directional detector<br />
  14. 14. Running behaviour recognition<br />
  15. 15. Time-based loitering behaviour recognition<br />
  16. 16. Split target /abandon luggage detection<br />
  17. 17. Group people tracking<br />
  18. 18. Tampering and malfunction detectors<br />Loss of signal<br />Obstruction<br />Out of focus and lens dusting<br />Blackout and overexposure <br />AE failure<br />Lightingfailure<br />
  19. 19. Digital image stabiliser (antishaker)<br />Eliminates video shakingcaused by wind and industrial vibrations <br />Essential for analytics performance<br />Differentiates the camera movementsfrom scene background/foreground movements<br />
  20. 20. Video analytics components<br />
  21. 21. Object tracker complexity<br />complexity<br />
  22. 22. Dynamic texture of the real world<br />
  23. 23. Dynamic texture modelling<br />OBJECT<br />HAAR FEATURES<br />BACKGROUND<br />4D-pyramid<br />Featureprobability cloud<br />α-channel (mask) for each object<br />
  24. 24. People group tracking (Q4 2010)<br />Feature cloud enablesobject tracking under partial visibility<br />Z-buffer to identify object occlusions<br />
  25. 25. Long range intrusion detectionusing directional tripwire<br />Unlimited numberof tripwires<br />Metadata includetripwire number<br />Detection ofunidirectional or bidirectionalcrossing<br />
  26. 26. Rule based behaviour recognitionEach zone is configured independently<br />Zone entrance<br />Zone exist<br />Zone loitering:Staying overpredefined period of time<br />Zone running:<br />Exceeding a predefined speed<br />Directional move within zone<br />
  27. 27. Metadata sent over IP network / ONVIF<br />Event type, data and time<br />Zone or tripwire number<br />2D object feature:<br />Position, size, area, speed<br />Real 3D features<br />Estimated from 2D featuresusing calibration data<br />JPEGframe image withobject trajectory annotation<br />
  28. 28. Videoanalytics calibration<br />Two human figures define scale & angle<br />Drag’n’drop calibration<br />Tracking region<br />2D to 3D coordinate transform<br />
  29. 29. Video analytics parameters<br />Service detectors<br />Antishaker<br />Object tracker<br />Contrast sensitivity<br />Special sensitivity<br />Min. stabilisation time<br />Object filters<br />Maximum object speed<br />Min and max areas<br />1<br />2<br />3<br />4<br />
  30. 30. Video analytics evaluation<br />Methods and results<br />
  31. 31. Video analytics public tests<br />
  32. 32. Sterile Zone Performance<br />38 hours, PAL (720 x 576 x 25 fps), M-JPEG, 40 Mbps<br />Number of true positive alarms: a = 432<br />False positivesalarms (typeI error): b =2<br />False negativesalarms (typeII error): с= 0<br />
  33. 33. Object detection range<br />
  34. 34. Range doubled with HD analytics<br />15-25 m<br />20-30 m<br />25-45 m<br />
  35. 35. Maximum response time<br />People walking and running<br />2 seconds<br />People moving slowly(e.g. crawling)<br />10 seconds<br />
  36. 36. Causes of false negatives(simple motion detectors)<br />Unstable background decreasessensitivity of an adaptive detector<br />DYNAMIC TEXTURE MODELING ALGORITHMSENABLE ROBUST OBJECT DETECTION IN A CHALENGING ENVIROMENT<br />
  37. 37. Causes of false positives(basic motion detectors)<br />Variable lighting<br />Shadows from moving clouds and sun<br />Moving trees, bushes and water<br />Camera shaking<br />Animals, birds and insects<br />Object trajectory split and double detection<br />Snow, rain, fog<br />
  38. 38. Examples of false positives(simple motion detectors)<br />BIRD<br />RABBIT<br />INSECT<br />CAMERA SHAKING<br />VIDEO ANALYTICS PREVENTS FALSE ALARMS CAUSED BY THESE FACTORS<br />
  39. 39. Object trackingwhilst tree shadows moving<br />
  40. 40. Performance estimation by3D security modeling<br />3D modeling<br />building infrastructure<br />control zones of camerasand third-party detectors<br />treats (in space-time)<br />Estimation of detection probabilities under variable external conditions<br />day/night, fog, snow<br />Video presentation<br />ORIGINAL BUILDING<br />3D MODEL OF BUILDNG<br />
  41. 41. Hardware reference designs<br />Multifunctional video services and HD cameraswith embedded analytics<br />
  42. 42. System-on-chip video analytics<br />Videofilters<br />Linux<br />Video<br />analytics<br />HD H.264 codec<br />1080p<br />
  43. 43. Dual channel video analytics encoder<br />3/17/2010<br />43<br />ANALOG + IPHYBRID TECHNOLOGY<br />Two analogue inputs (BNC)<br />Two managed outputs (BNC)and digital video over IP<br />H.264 &MJPEG encoding<br />Embedded video & audio analytics<br />POE+and backup power<br />ONVIF 1.01 support<br />- 40⁰...+50⁰С<br />Lightning guard<br />
  44. 44. HD video analytics camera<br />
  45. 45. MJPEG vsH.264 compression<br />DATAFLOW, MBPS<br />RESOLUTION<br />
  46. 46. Unique selling position<br />Fully embedded (DSP) implementation<br />Real-time processing of uncompressed video<br />HD/Megapixel resolution<br />Highly scalable<br />Unmatched performance in sever environment<br />dynamic texture engine<br />End-user hardware i-LIDS certification<br />on schedule April 2010<br />Wide interoperability<br />ONVIFcompliance<br />
  47. 47. Future of video surveillance<br />Multiple camera tracking using 3D model<br />
  48. 48. Segmentation problemand object occlusions<br />‘Single camera’video analytics<br />‘Multiple camera’video analytics<br />A<br />B<br />C<br />A<br />
  49. 49. i-LIDS multiple camera tracking scenario<br />2<br />3<br />4<br />
  50. 50. 17/03/2010<br /><br />50<br />Video analytics + 3D modeling<br />3D model of a buildingand camera controlzones<br />1<br />2<br />Камера 2<br />Камера 1<br />
  52. 52. 3D trajectory reconstructed frommultiple video sources<br />