video surveillance using distance maps january 2006 theo schouten harco kuppens egon van den broek

15
Video Surveillance using Distance Maps January 2006 Theo Schouten Harco Kuppens Egon van den Broek

Post on 15-Jan-2016

215 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Video Surveillance using Distance Maps January 2006 Theo Schouten Harco Kuppens Egon van den Broek

Video Surveillance using Distance Maps

January 2006

Theo SchoutenHarco Kuppens

Egon van den Broek

Page 2: Video Surveillance using Distance Maps January 2006 Theo Schouten Harco Kuppens Egon van den Broek

Video Surveillance using Distance Maps

Video Surveillance

• fast growing sector in security market• fundamental issues and challenges

– interpretation, generality, automation, efficiency, robustness, trade off, performance evaluation, multiple camera and data fusion, feature selection and integration(Amer and Regazzoni)

• efficiency (real-time) and robustness• single camera, top view moving+stationary objects• detect objects, measure distances + motion• abstract from color to binary conversion

– model imperfections: changing illumination, shadows, video noise

Page 3: Video Surveillance using Distance Maps January 2006 Theo Schouten Harco Kuppens Egon van den Broek

Video Surveillance using Distance Maps

Operational environment

• virtual, dynamic robot navigation environment– binary frames with moving+stationary objects

using Macromedia Flash

• noise model– border object pixel:

p1% ->background pixel

– random chosenbackground neighbor:p2%->object pixel

– each pixelp3% -> inverse

60 320x240 frames50%-50%-5% noise

Page 4: Video Surveillance using Distance Maps January 2006 Theo Schouten Harco Kuppens Egon van den Broek

Video Surveillance using Distance Maps

Large sequence

120 640x480 frames once spontaneous movement10%-10%-1% noise of stationary objectchanging number of once a collision moving objects

Page 5: Video Surveillance using Distance Maps January 2006 Theo Schouten Harco Kuppens Egon van den Broek

Video Surveillance using Distance Maps

Fast Exact Euclidean Distance (FEED) Maps

• D(p) = if (p O) then 0 else each q O

feeds its ED to each p:D(p) = min ( D(p), ED(q,p))

(10-20 ms, factor 2 slower than chamfer 3,4)

• ED map stationary objects only:– loop over border moving object: ED to stationary objects

• ED stat+moving=min(ED stat,ED moving)

(0.5-1 ms, factor 2 faster than chamfer 3,4)– input to ”robot” objects

border pixels bisection lines precalculate ED

Page 6: Video Surveillance using Distance Maps January 2006 Theo Schouten Harco Kuppens Egon van den Broek

Video Surveillance using Distance Maps

Real-time and exact motion detection

• initialization: n (5) frames to locate stationary pixels• per frame:

– determine pixels of stationary and moving objects

– check for a movement of stationary objects

– locate moving objects

– calculate distances

– generate output (application dependent)

• list of tracked (frame-to-frame) objects+distances

• graphical display of objects+distance

• for 1 “robot”: ED map of stationary and other moving objects

Page 7: Video Surveillance using Distance Maps January 2006 Theo Schouten Harco Kuppens Egon van den Broek

Video Surveillance using Distance Maps

Design guidelines for speed

• pre-calculate– data structures depending only on stationary obj.

• avoid data movement– keep track of added moving object data– reinitialize only changed parts

• minimize loops and test– combine logically distinct program parts– split a logical function over program parts

• use the right level of abstraction– stationary: pixels; moving: objects

Page 8: Video Surveillance using Distance Maps January 2006 Theo Schouten Harco Kuppens Egon van den Broek

Video Surveillance using Distance Maps

Output display: objects and distances

60 320x240 frames50%-50%-5% noise

Page 9: Video Surveillance using Distance Maps January 2006 Theo Schouten Harco Kuppens Egon van den Broek

Video Surveillance using Distance Maps

Output display: ED map for 1 object

60 320x240 frames50%-50%-5% noise

Page 10: Video Surveillance using Distance Maps January 2006 Theo Schouten Harco Kuppens Egon van den Broek

Video Surveillance using Distance Maps

Timing

120 640x480 frames

50%-50%-5% noise

AMD

1666 MHz

Intel M725

1600 MHz

Initiali-zation

processing 46.60 ms 27.79 ms

display

generation

10.33 ms 14.54 ms

Per frame

processing 4.94 ms 3.21 ms

display

generation

1.40 ms 1.99 ms

Page 11: Video Surveillance using Distance Maps January 2006 Theo Schouten Harco Kuppens Egon van den Broek

Video Surveillance using Distance Maps

Details: locating stationary object pixels

• moving objects should move sufficiently fast:– no overlap in at least 2 frames

• if not:– program keeps running– but too often in initialization phase

• further strategies:– adapt number of initialization frames– more elaborate statistical processing– towards object detection

Page 12: Video Surveillance using Distance Maps January 2006 Theo Schouten Harco Kuppens Egon van den Broek

Video Surveillance using Distance Maps

Details: minimum movement stationary objects

• red: disappeared stationary object pixels• 22, 54 and 73 (least noise sequences: 36,99 and 92)• maximum red pixels due to noise: 2 (0)• able to detect very small movements robustly• dependent on “imperfection and noise” model:

– not direction dependent, no form change• strategies: skip frames, appearing object pixels, etc.

Page 13: Video Surveillance using Distance Maps January 2006 Theo Schouten Harco Kuppens Egon van den Broek

Video Surveillance using Distance Maps

Details: minimum size moving objects

• “hole” noise objects:

– removed by a simple, fast method

– in theory pathological cases where this will fail

• other noise objects: removed by threshold on size

– contour size: noise maximal 9, minimal object: 42

– moving objects can be factor 3 smaller

part input frame red= moving color: border moving

Page 14: Video Surveillance using Distance Maps January 2006 Theo Schouten Harco Kuppens Egon van den Broek

Video Surveillance using Distance Maps

Conclusion

• real-time, robust object, distance and motion detection– well defined environment, with limitations

– using distance maps generated by FEED

– providing output for surveillance purposes

• design guidelines to achieve our results• discussed 3 restrictions on content of frames• pointers to further research

– reduce the restrictions

– enlarge variability of environment• simulated environment with other “noise” models• real video camera input

Page 15: Video Surveillance using Distance Maps January 2006 Theo Schouten Harco Kuppens Egon van den Broek

Video Surveillance using Distance Maps

The End