Savvy Vision Processor .


42 views
Uploaded on:
Category: Product / Service
Description
Intelligent Vision Processor. John Morris Computer Science/ Electrical & Computer Engineering, The University of Auckland. “Iolanthe II” rounds Channel Island - Auckland-Tauranga Race, 2007. Intelligent Vision Processor. Applications  Robot Navigation
Transcripts
Slide 1

Shrewd Vision Processor John Morris Computer Science/Electrical & Computer Engineering, The University of Auckland "Iolanthe II" rounds Channel Island - Auckland-Tauranga Race, 2007

Slide 2

Intelligent Vision Processor Applications  Robot Navigation  Collision shirking – self-governing vehicles  Maneuvering in element situations  Biometrics Face acknowledgment  Tracking people  Films  Markerless movement following  Security  Intelligent risk identification  Civil Engineering  Materials Science  Archeology

Slide 3

Background

Slide 4

Intelligent Vision Our vision framework is uncommon Capabilities right now surpass those of any single processor Our brains Operates on a moderate \'clock\': kHz area Massively parallel >10 10 neurons can figure in parallel Vision framework (eyes) can abuse this parallelism ~3 x 10 6 sensor components (poles and cones) in human retina

Slide 5

Intelligent Vision Matching and acknowledgment Artificial insight frameworks are at present not in the race! For instance Face acknowledgment We can perceive faces From shifting points Under outrageous lighting conditions With or without glasses, facial hair, wraps, cosmetics, and so on With skin tone changes, eg sunburn Games We can strike balls going at > 100 km/h and Direct that ball with high exactness

Slide 6

Human vision Uses a moderately moderate , however enormously parallel processor ( our brains ) Able to perform undertakings At rates and With precision past abilities of best in class fake frameworks

Slide 7

Intelligent Artificial Vision High execution processor Too moderate for high determination (Mpixel+) picture continuously (~30 outlines every second) Useful vision frameworks Must have the capacity to Produce 3D scene models Update scene models rapidly Immediate objective: 20-30Hz to copy human capacities Long term objective: >30 Hz to give upgraded abilities Produce precise scene models

Slide 8

Intelligent Artificial Vision Use human cerebrum as the central model We know it works superior to an ordinary processor! We require

Slide 9

Lens Eyeball Retina Human Vision Systems Higher request creatures all utilization binocular vision frameworks Permits estimation of separation to a protest Vital for some survival undertakings Hunting Avoiding peril Fighting predators Distance (or profundity) figured by triangulation P\'\' P\' P\'- P\'\' is the difference It increments as P comes nearer

Slide 10

Human Vision Systems Higher request creatures all utilization binocular vision frameworks Permits estimation of separation to a question Vital for some survival assignments Hunting Avoiding risk Fighting predators Distance (or profundity) registered by triangulation P\'\' P\' P\'- P\'\' is the uniqueness Increases as P comes nearer

Slide 11

P Optical hub P\'\' P\' Fixation point P F P\'\' P\' Artificial Vision Evolution took a large number of years to enhance vision Don\'t overlook those lessons ! Binocular vision works Verging optics Human eyes are known to swivel to "focus" on a protest of intrigue

Slide 12

Real versus Ideal Systems Real focal points misshape pictures Distortion must be evacuated for high accuracy work! Simple yet Conventional strategy utilizes iterative arrangement Slow! Quicker approach required for ongoing work Image of a rectangular matrix with a genuine focal point

Slide 13

Why Stereo? Run discoverers give profundity data straightforwardly SONAR Simple Not extremely exact (long l ) Beam spread  Low spatial determination Lasers Precise Low disparity  High spatial determination Requires genuinely advanced gadgets Nothing excessively difficult in 2008 Why utilize a roundabout estimation when coordinate ones are accessible?

Slide 14

Why Stereo? Uninvolved Suitable for thick situations Sensors don\'t meddle with each other Wide territory scope Multiple covering perspectives reachable without obstruction Wide region 3D information can be obtained at high rates 3D information helps unambiguous acknowledgment 3 rd measurement gives extra separation Textureless areas cause issues yet Active enlightenment can resolve these Active examples can utilize IR (undetectable, eye-safe) light

Slide 15

Artificial Vision Challenges

Slide 16

Artificial Vision - Challenges High processor control Match parallel capacities of human cerebrum Distortion expulsion Real focal points dependably demonstrate some twisting Depth precision Evolution learnt about skirting optics a large number of years prior! Proficient coordinating Good corresondence calculations

Slide 17

Artificial Vision Simple stereo frameworks are being delivered Point Gray, and so forth All utilization authoritative design Parallel tomahawks, coplanar picture planes Computationally more straightforward High execution processor doesn\'t have room schedule-wise to manage the additional computational multifaceted nature of skirting optics Point Gray Research Trinocular vision framework

Slide 18

Artificial System Requirements Highly Parallel Computation Calculations are not mind boggling but rather There are a great deal of them in megapixel+ ( >10 6 ) pictures! High Resolution Images Depth is ascertained from the divergence If it\'s exclusive a couple of pixels, then profundity precision is low Basic condition (standard design just!) Baseline Focal Length Depth, z = b f d p Pixel measure Disparity

Slide 19

Artificial System Requirements Depth determination is basic! A cricket * player can get a 100mm ball going at 100km/h High Resolution Images Needed Disparities are extensive quantities of pixels Small profundity varieties can be measured yet High determination pictures increment the interest for handling power! *Strange diversion played in previous British provinces in which a batsmen safeguards 3 little sticks in the focal point of a vast field against a bowler who tries to thump them down!

Slide 20

Artificial System Requirements Conventional processors don\'t have adequate preparing power however Moore\'s Law says Wait year and a half and the power will have multiplied yet The progressions that give you double the power additionally give your twice the same number of pixels consecutively and four times the same number of in a picture! Concentrated very parallel equipment is the main arrangement!

Slide 21

Processing Power Solution

Slide 22

FPGA Hardware FPGA = Field Programmable Gate Array "Delicate" equipment Connections and rationale capacities are "customized" similarly as a customary von Neuman processor Creating another circuit is about as troublesome as composing a program! High request parallelism is simple Replicate the circuit n times As simple as composing a for circle!

Slide 23

FPGA Hardware FPGA = Field Programmable Gate Array "Circuit" is put away in static RAM cells Changed as effortlessly as reloading another program

Slide 24

FPGA Hardware Why is programmability essential? then again Why not outline a custom ASIC? Optical frameworks don\'t have the adaptability of a human eye Lenses created from inflexible materials Not conceivable to make a \'one framework fits all\' framework Optical arrangements must be intended for every application Field of view Resolution required Physical limitations … Processing equipment must be adjusted to the optical setup If we plan an ASIC, it will work for one application!!

Slide 25

Correspondence or Matching

Slide 26

Stereo Correspondence Can you discover all the coordinating focuses in these two pictures? "Obviously! It\'s simple!" The best PC coordinating calculations get 5% or a greater amount of the focuses totally off-base! … and set aside a long opportunity to do it! They\'re not possibility for constant frameworks!!

Slide 27

Stereo Correspondence High execution coordinating calculations are worldwide in nature Optimize over substantial picture areas utilizing vitality minimization plans Global calculations are inalienably moderate Iterate many times over little districts to discover ideal arrangements

Slide 28

Correspondence Algorithms Good coordinating execution, worldwide, low speed Graph-cut, conviction proliferation, … High speed , straightforward, nearby, high parallelism, most minimal execution Correlation High speed , direct intricacy, parallel, medium execution Dynamic programming calculations

Slide 29

Depth Accuracy

Slide 30

Points thusly have similar difference Stereo Configuration Canonical setup – Two cameras with parallel optical tomahawks Rays are drawn through every pixel in the picture Ray crossing points speak to focuses imaged onto the focal point of every pixel Depth determination yet To get profundity data, a point must be seen by both cameras, ie it must be in the Common Field of View

Slide 31

Stereo Camera Configuration Now, consider a protest of degree, a To be totally measured, it must lie in the Common Field of View yet put it as near the camera as you can so you can acquire the best precision, say at D Now increment b to expand the exactness at D But you should build D so that the question remains inside the CFoV! Itemized investigation prompts an ideal estimation of b  an a D b a

Slide 32

Increasing the gauge Increasing the benchmark diminishes execution!! % great matches Images: "passageway" set (beam followed) Matching calculations: P2P, SAD Baseline, b

Slide 33

Increasing the standard Examine the dissemination of blunders Increasing the gauge diminishes execution!! Standard Deviation Images: "hallway" set (beam followed) Matching calculations: P2P, SAD Baseline, b

Slide 34

Increased Baseline  Decreased Performance Statistical Higher uniqueness go expanded likelihood of coordinating mistakenly - you\'ve essentially got more options! Viewpoint Scene articles are not fronto-planar Angled to camera tomahawks subtend distinctive quantities of pixels in L and R pictures Scattering Perfect dispersing (Lambertian) surface suspicion OK at little rakish contrasts expanding disappointment at higher edges Occlusions Number of shrouded locales increments as precise distinction increments expanding number of "monocular" focuses for which there is no 3D data!

Slide 35

Evolut

Recommended
View more...