340 likes | 469 Views
MULTI-OMNIDIRECTIONAL CAMERAS FOR AMBIENT INTELLIGENCE. Bar ış Evrim Demiröz, Albert Ali Salah, Lale Akarun. 19.04.2012 Sinyal İşleme ve İletişim Uygulamaları Kurultayı. Omnidirectional Camera. Camera with large field of view ~180° >180° ~360 °. Omnidirectional Camera.
E N D
MULTI-OMNIDIRECTIONALCAMERAS FOR AMBIENTINTELLIGENCE Barış Evrim Demiröz, Albert Ali Salah, Lale Akarun 19.04.2012 Sinyal İşleme ve İletişim Uygulamaları Kurultayı
Omnidirectional Camera • Camera with large field of view • ~180° • >180° • ~360 ° Multi-Omnidirectional Cameras for Ambient Intelligence
Omnidirectional Camera • Camera with large field of view • ~180° • >180° • ~360 ° Multi-Omnidirectional Cameras for Ambient Intelligence
Omnidirectional Cameras • Multiple images • PTZ, multiple cameras • Dioptric (lens) • Fisheye • Catadioptric (lens + mirror) Multi-Omnidirectional Cameras for Ambient Intelligence
Omnidirectional Cameras • Throwable Panoramic Ball Camera Photo: Jonas Pfeil Multi-Omnidirectional Cameras for Ambient Intelligence
Omnidirectional Cameras • Pan Tilt Zoom (PTZ) camera Photo: Wikimedia Commons Multi-Omnidirectional Cameras for Ambient Intelligence
Omnidirectional Cameras • Fisheye lens Photo: Wikimedia Commons Multi-Omnidirectional Cameras for Ambient Intelligence
Omnidirectional Cameras • Catadioptric (lens + mirror) B. Micusik and T. Pajdla, "Estimation of omnidirectional camera model from epipolar geometry" 2003 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2003. Proceedings., pp. I-485-I-490, 2003. Multi-Omnidirectional Cameras for Ambient Intelligence
Catadioptric/Dioptric cameras • Pros/cons • 1 image • Low resolution near boundaries • Ideal for small spaces, surveillance Title of the presentation
Dataset Multi-Omnidirectional Cameras for Ambient Intelligence
Sample from dataset Multi-Omnidirectional Cameras for Ambient Intelligence
The room Pencereler Masa 1 YanKamera Masa 2 Sandalye 1 Sandalye 2 Üst Kamera Lavabo Kapı Multi-Omnidirectional Cameras for Ambient Intelligence
Cameras used • Oncam IPC, IPT • 5 Megapixel • No moving parts Multi-Omnidirectional Cameras for Ambient Intelligence
Scenarios performed • Scenario #1 • Single person • 6 Actions • Sitting • Walking • Drinking • Washing hands • Fainting • Opening/Closing door • 5 performers (10 videos total) Multi-Omnidirectional Cameras for Ambient Intelligence
Scenarios performed • Scenario #2 • 3 people • 5 Actions • Sitting • Walking • Standing • Handshaking • Interested in object • 5 performers (36 videos intotal by exchanging roles) Multi-Omnidirectional Cameras for Ambient Intelligence
Video properties • 640x480 frame size • Bandwidth limitation • ~8 fps • MJPEG format Multi-Omnidirectional Cameras for Ambient Intelligence
Annotations C • Annotation tool usedvatic* • Annotations • Bounding box of subject • State (action) of subject A B * C. Vondrick, D. Ramanan, and D. Patterson, “Efficiently Scaling Up Video Annotation with Crowdsourced Marketplaces,” ECCV 2010, pp. 610–623, 2010. Multi-Omnidirectional Cameras for Ambient Intelligence
Illumination conditions • Intensity standard deviation Multi-Omnidirectional Cameras for Ambient Intelligence
Occlusion • Pseudo occlusion: Multi-Omnidirectional Cameras for Ambient Intelligence
Occlusion Multi-Omnidirectional Cameras for Ambient Intelligence
How to evaluate tracker performance? • Multiple Object Tracking Precision (MOTP)* • Multiple Object Tracking Accuracy (MOTA)* * K. Bernardin and R. Stiefelhagen, “Evaluating Multiple Object Tracking Performance: The CLEAR MOT Metrics,” EURASIP Journal on Image and Video Processing, vol. 2008, pp. 1–10, 2008. Multi-Omnidirectional Cameras for Ambient Intelligence
How to evaluate tracker performance? o: object h: hypothesis d: distance btw o: and h c: match count m: misses fp: false positives mme: mismatches t: frame no k: video no i: object no Multi-Omnidirectional Cameras for Ambient Intelligence
Released to public • Videos • Annotations • Performance Measurer (in Java) • Camera calibrations • Source codes Multi-Omnidirectional Cameras for Ambient Intelligence
Base tracker Multi-Omnidirectional Cameras for Ambient Intelligence
Outline System Components Multi-Omnidirectional Cameras for Ambient Intelligence
Background/Foreground Segmentation • Mixture of Gaussians • Shadow removal Multi-Omnidirectional Cameras for Ambient Intelligence
Threshold the areas filter small contours Blob formation • Morphological close • maintain connectivity × × Multi-Omnidirectional Cameras for Ambient Intelligence
FAST*corner detection Feature Point Detection andDescriptor Extraction • BRIEF features • extended for opponent color space • *Features from Accelerated Segment Test Multi-Omnidirectional Cameras for Ambient Intelligence
Blob Matching and Tracking Frame Blob Blob Blob … Keypoint Keypoint Keypoint Keypoint Keypoint Keypoint … … … Keypoint Keypoint Keypoint Multi-Omnidirectional Cameras for Ambient Intelligence
Blob Matching and Tracking Previous n frames Frame NEW FRAME Frame Frame … Blob Blob Blob Blob Blob Blob Blob Blob Blob Blob Blob Blob … … … … Keypoint Keypoint Keypoint Keypoint Keypoint Keypoint Keypoint Keypoint Keypoint Keypoint Keypoint Keypoint Keypoint Keypoint Keypoint Keypoint Keypoint Keypoint Keypoint Keypoint Keypoint Keypoint Keypoint Keypoint … … … … … … … … … … … … Keypoint Keypoint Keypoint Keypoint Keypoint Keypoint Keypoint Keypoint Keypoint Keypoint Keypoint Keypoint Multi-Omnidirectional Cameras for Ambient Intelligence
Blob Matching and Tracking • Optimal assignment problem • Hungarian algorithm Multi-Omnidirectional Cameras for Ambient Intelligence
Results Multi-Omnidirectional Cameras for Ambient Intelligence