new application areas made possible by high speed vision workshops/2011 workshop... · new...
TRANSCRIPT
New Application Areas Made Possible by High Speed Vision
Masatoshi IshikawaUniversity of Tokyo
http://www.k2.t.u-tokyo.ac.jp/
M. Ishikawa http://www.k2.t.u-tokyo.ac.jp/
Decomposition Problem of System Tasks
Hierarchical Parallel System
2
Real World
Real Time Parallel Processing
Enough Bandwidth without Bottleneck
Hierarchical Parallel Distributed System with High Speed Vision
judgmenttask
switchtask
switch
patternrecognition
trajectorygeneration
obstacleavoidance
stereovision
tactilevision
visualservo
imagefeature
tactilemotortrajectorytrajectory
vision visiontactilemotormotormotor
M. Ishikawa http://www.k2.t.u-tokyo.ac.jp/
Task Decomposition
3
sensors recognition planning control actuators
Sequential Decomposition
Sufficient Condition of Parallel Decomposition:
Orthogonality among Tasks
(Time and/or Space)
Orthogonal Decomposition
sensors
Tracking
Reaching
Preshaping
Grasping
actuators
Parallel DecompositionM. Ishikawa http://www.k2.t.u-tokyo.ac.jp/
Dynamics Matching
4
EnvironmentProcessing
Sensors
Actuators
Dynamics
Frequency
Gai
n
Dynamics
Frequency
Gai
n
Artificial elements, i.e. sensors, actuators, and processing modulesshould be designed for realizing enough dynamics matched with thedynamics of object and environment.
→ Performance up of sensors, actuators, and processing modules
→ Keep the sampling theorem
Dynamics
Frequency
Gai
n
Dynamics
Frequency
Gai
n
M. Ishikawa http://www.k2.t.u-tokyo.ac.jp/ 5
Visual Feedback
Imaging device
Image processing
Lighting
High Speed Vision
System
Fusion Network
Recognition / Understanding
Object
Variable Flame Rate up to 1,000 FPS
PD Array PE Array
PD Array PE Array
Standard InterfaceArchitecture
Camera Link et al.
High Speed Image Processing
fram
e ra
te
256×256 512×512 1M
1kHz(1ms)
100Hz(10ms)
(pixels)
resolution4M
30Hz(33ms)
High SpeedVision
Image Processingwith Scanning
Fully Parallel
Fully Parallel Architecture
Column Parallel Architecture
M. Ishikawa http://www.k2.t.u-tokyo.ac.jp/
Architecture of Vision Chip
6
Output
OutputCircuit Decoder
ProcessingElement
PhotoDetector
Clock
・・・
Instruction
ALU
D-Latch
D-Latch
D-Latch
I0-I4
4-neighbors
sensoroutput 0
I0-I4W_EN
Z_EN
B_EN
A_EN
Local Memory
Memory-mapped I/O400 Trs
Processing Element
ALU67.4μm
67.4μ
m
32×48 PEwith lens
col_sum0
0
0
0
soutcol_decoder
row
_dec
oder
PE
col_inst
row
_ins
t
HALFADDER
CS
D-FFMU
X
col(x )
D FF
MU
X
clka
g(x-1,y)PD
r(x,y)
f(x,y)
mout
crop
row(y)
r(x- , y)
clkbsel
load
g(x,y)
1
-
g(x+1,y)
g(x,y-1)g(x,y+1)
Summation
Tracking
84Trs.
Vision Chip with Massively Parallel Processing
Target Tracking Vision Chip
Developed in cooperation with Seiko NPC Corp.
M. Ishikawa http://www.k2.t.u-tokyo.ac.jp/
Imager + PE Array
7
Developed in cooperation with Hamamatsu Photonics K.K.
PD array
128×128pixels
PE array
128×128pixels
ColumnParallelImage
Transfer
ADC
8bit×128
ImageInput
SummationCircuit
Controller
Control Signals Instructions
PixelData In/Out
Image Feature Extraction
Output(Image Feature)
1,280×1024 pixels 500 fps1,280× 512 pixels 1,000 fpsInterface: Ether net, USB, etc.
Camera Link Full (650MB/s)
Column Parallel Vision with Massively Parallel Processing
Imager + Reconfigurable PE Array
Camera IF
Camera Link
FPGA1
FPGA2
CPU
RS232C
RAM RAM
DIO
LAN USB
VGA
M. Ishikawa http://www.k2.t.u-tokyo.ac.jp/
High Frame Rate Vision
8
High Frame Rate Makes Algorithms Simple! Sampling Theorem in Image Space
Camera
Object
x= m10 / m00 , y= m01 / m00
Center of Gravity- -
Position DetectionSAD
Sum
Model Image M
Input Image I
AD
ModelParameter pDraw
Modification
Model Based Shape Recognition
Field of View
Actuator + Encoder
(x0, y0)
Input Image
∩ =
Window Target
Sk iWk-1 iCk
Input Image
∩ =
Window Target
Sk+1 iWk iCk+1
Target Tracking
I[n] W[n-1]
∩
Carry Out Target Tracking Algorithm in Sequence of Object
2 3
1
W[n]Multitarget Tracking
2 3
1
Active Vision
Fixed Type
M. Ishikawa http://www.k2.t.u-tokyo.ac.jp/
Applications
9
RoboticsHigh Speed Intelligent Robots
High-Throughput Product Lines
Human InterfaceInput Devices for Computers / Games
PD array
XY stageSPE
Bio / MedicalRemote Medical CareSurgery Aid
Nano / Micro MachinePosition Control Object Detection
e-BooksHigh Speed Scanning
Other Applications where VIDEO rate vision is slow.
Inspection
VehicleIntelligent VehicleITS / Air Craft
SecurityHuman TrackingMotion Detection
MediaObject Tracking Video Control
M. Ishikawa http://www.k2.t.u-tokyo.ac.jp/
High Speed Gesture UI 1
10
32×48 pixels1,000 fps(2002)Sub-pixel resolution has been performed.
Game Control by Gesture Mouse Operation by Gesture
Gesture Recognition
16×16 pixels1,000 fpsActive Vision(1994)
Perfect Understanding of Human Behavior
Wearable Computer Ubiquitous Computing
Game / TV ControlMulti Modal Interface
Non-Touch, Cable Free, High Speed, 3D Input
M. Ishikawa http://www.k2.t.u-tokyo.ac.jp/
High Speed Gesture UI 2
11
immersive and proprioceptive interface
Computer or Game
High Speed Vision
High Speed and Low Latency Gesture Recognition
Type in the Air
1,000 fps30ms Total Latency
High Speed Vision
154 fpsFinger Gesture
ClickingTyping3D Input Infrared LED
M. Ishikawa http://www.k2.t.u-tokyo.ac.jp/
Multi Target Tracking
12
1,000 Objects Tracking at 1,000 FPS
Book Flipping Scanning
プロジェクタ
カメラ
測定領域
Template Matching for 1,000 Objects
Structured Light3D Shape MeasurementImage Reconstruction
LED LaserCamera
M. Ishikawa http://www.k2.t.u-tokyo.ac.jp/
Dynamic Imaging
13
High Speed Camera224×64 pixel, 1,000 fpsMosaicking
Throwing Camera
Inspection of Rotating Can Surface
Surface Print Inspection1,000 fps1280×512 pixelMosaicking
View of a Developed Surface
M. Ishikawa http://www.k2.t.u-tokyo.ac.jp/
Micro Visual Feedback
14
object
intelligent vision system (CPV)
•Microscopic Target Tracking•Visual Feedback Control of X-Y Stage•1ms Feedback Rate
・Microscope Surgery Aid・Biological Experiment / Inspection・Nano/micro Machine Control
XY stage
CCD camera
Small Object Tracking(Paramesium) (Ascidian Spermatozoa)
Elimination of Tremor(Surgery Robot)
M. Ishikawa http://www.k2.t.u-tokyo.ac.jp/
Variable Focus Lens
15
High Speed Visual Feedback
High Speed Variable Focus Lens
Autofocus-Scanning Microscope
Dynamorph Lens
Variable Focus Lens
3D scanning TrackingObject Parallel
M. Ishikawa http://www.k2.t.u-tokyo.ac.jp/
High Speed Adaptive Optics
16
35
Object LensDo=40
Field LensDf=40
Collimator Dc=40
8060
fo=80fc=60
ff =100
35
Camera30
Pupil
All In-Focus Image by Using Variable Focus Lens
Saccade Mirror: High-speed Gaze Control System
Focus Scan at 8,000 Hz → 1,000 fps All In-Focus Image
Dynamorph Lens →
M. Ishikawa http://www.k2.t.u-tokyo.ac.jp/
High Speed Robotics
17
Robot is a dynamic information fusion system which consists of sensor, processing, and actuator.
Robot should be:→ more intelligent
than humans→ too fast to see
Intelligent Robots
IndustrialRobots
Human Intelligence Our Target
Speedslow fast
Fle
xib
ility
General Purpose
SpecialPurpose
Human is not a final target of robotics.
Challenge to the Theoretical Limits (Sampling Theorem)
Challenge to the Intelligence (Dynamics of Intelligence)
Elimination of Bottleneck (Development of Devices)
Static Control→ Dynamic Control (High Performance)
1M. Ishikawa http://www.k2.t.u-tokyo.ac.jp/
1ms Sensor Fusion System
18
High-speed ArmMax speed: 5m/sMax acceleration: 6g
High-speed Hand3 Fingers, 8DOF, 800g, 180deg/0,1s
TactileSensor
Other Sensors
DSP
Host
DSP DSP DSP
Parallel Processing System Real Time Processing 1ms Feedback Rate
DSP
High Speed Vision Resolution: 128x128, 8bitProcessing rate: 1KHz
CPV
CPV
M. Ishikawa http://www.k2.t.u-tokyo.ac.jp/
High Speed Robot Arm
19
Batting Robot
Collision AvoidanceBatting-Throwing Robots
Swing ModeLocal Feedback
Hitting Mode
Target Tracking
World Coordination
Local Coordination
M. Ishikawa http://www.k2.t.u-tokyo.ac.jp/
Dynamic Catching
20
High-speed Hand
Active Vision
Target
Active Vision
Dynamic Catching Using High Speed Vision
Raw Egg Catching
M. Ishikawa http://www.k2.t.u-tokyo.ac.jp/
Dynamic Manipulation 1
21
v1 v2
0v ω0
V1, V2
V0, ω0
P, θ
before throwing after throwing catching
n
Regrasping
Dribbling
M. Ishikawa http://www.k2.t.u-tokyo.ac.jp/
Dynamic Manipulation 2
22
Rope Knotting
Cloth Folding
Catch in the Air
Manipulation of Flexible Materials
High speed motion with visual feedback makes tasks so Simple.
M. Ishikawa http://www.k2.t.u-tokyo.ac.jp/
Application Systems
23
Human Interface
High Speed Robots
FA and Inspection
Bio/Medicine
Media/ArtVehicle, ITS
Human TrackingNetwork Sensing
Observationof Cell
Gesture UI
ITS, Auto Driving, Collision Avoidance
Interactive Art
High Speed Tracking Imaging
3D DisplayHaptic Rader
Target Tracking Vision Chip
Special Purpose Chip with Lens
High Speed Vision
Pixel Parallel Programmable Vision Chip
Pixel ParallelSpecial Purpose Vision Chip
Pixel Parallel ProgrammableBoard
Column Parallel Vision System
PD Separated Vision System
Variable Focus Lens
Lens and System
High Speed HandSecurity
Removal of Tremor
Synchronized Video
Material Handling
Catch in the Air
Regrasping
Dribbling
Pen Rotation
BattingThrowing
Collision Avoidance
Hig
h S
peed
Rob
ot
Han
ds
Surface Inspection
Multi Target Tracking andInspection
Book Flipping Scanning
Game Operation
3D Input
Mouse Operation
All In-Focus Image
Active Vision
3DMeasurement
MicroscopeInspection
Throwing Camera
High SpeedGesture UI
Smart LaserScanner (3D)
M. Ishikawa http://www.k2.t.u-tokyo.ac.jp/
Future of High Speed Vision
Smart Vision and System- Increase of Number of Sensors
- High Performance Computing and Network
- Progress in Semiconductor Technology
- Increase Space Density and Time Density
- Huge Number of Distributed Network Nodes
-Adequate Sampling and Processing Speed
Faster and More Intelligent !
In the next decade
24
PE PE
PEPEPE
: Processing Element: Sensor
PE
DataProgram
Data
Program
Distributed Smart SensorVLSI Integration (SoC Sensor)
Network TechnologyParallel Processing
PE