eMorph: towards neuromorphic robotic vision · A PCI based high-fanout AER mapper with 2 GiB RAM...

1
eMorph: towards neuromorphic robotic vision Information technology has not yet delivered artificial systems that can compare with biology in reliably, robustly and efficiently extracting information from the often noisy and ambigu- ous real world, and interact- ing with the world by generating appropriate behaviours. Tracker Motion Sensor - measures contrast, spatial and temporal derivative and velocity of moving edges - comprises circuits that implement a model of “selective attention” for a smart readout eMorph is an EU funded project - ICT-FET 231467 Visual System Visual Primitives Technology eMorph will deliver a methodology for event-driven artificial vision on a robotic demonstrator , giving the chance to the robotic community to have a ready-to-use platform to directly interact and experience the power of neuromorphic tools. Visual system integration We developed a series of printed circuit boards to integrate on the iCub event driven sensors and dedicated processors. iCub navigation We developed an autonomous vehicle with 6 holonomic, omnidirectional wheels. Asynchronous, event-driven, space variant visual sensors Goal: to design asynchronous, data-driven, biologically inspired, vision sensors with non-uniform morphology, using analog VLSI neuromorphic circuits High level event-driven visual algorithms Goal: to develop a supporting data-driven asynchronous computational paradigm for machine-vision that is radically different from conventional image processing System integration Goal: to validate the system on the humanoid robot iCub fitted with the neuromorphic sensors, and controlled by the asynchronous machine-vision algorithms in real-time behavioral tasks Publications Bartolozzi, C. and Indiveri, G., Selective Attention in Multi-Chip Address-Event Systems, Sensors, 2009 Orchard, G., Bartolozzi, C. and Indiveri, G., Applying neuromorphic vision sensors to planetary landing tasks, IEEE Biomedical Circuits and Systems Conference, 2009. Hofstätter, M., Schön, P. and Posch, C., A SPARC-compatible general purpose Address-Event processor with 20-bit 10ns-resolution asynchronous sensor data interface in 0.18μm CMOS, IEEE International Symposium on Circuits and Systems, 2010. Posch, C., Matolin, D., Wohlgenannt, R., Hofstaetter, M., Schoen, P., Litzenberger, M., Bauer, D., and Garn, H., Live Demonstration: Asynchronous Time-Based Image Sensor (ATIS) Camera with Full-Custom AE Processor, IEEE International Symposium on Circuits and Systems, 2010. Awarded as best live demonstration Fasnacht, D.B. and Indiveri, G., A PCI based high-fanout AER mapper with 2 GiB RAM look-up table, 0.8 us latency and 66MHz output event-rate, Conference on Information Sciences and Systems (CISS 2011), 2011. Bartolozzi, C., Metta, G., Hofstaetter, M. and Indiveri, G., Event-driven vision for the iCub, Bio-Mimetic and Hybrid Approach to Robotics, Workshop at IEEE International Conference on Robotics and Automation, 2011. Bartolozzi, C., Mandloi, N.K. and Indiveri, G., Attentive motion sensor for mobile robotic applications, IEEE International Symposium on Circuits and Systems, 2011. Bartolozzi, C., Fasnacht, D.B., Rea, F., Hofstaetter, M., Clercq, C., Metta, G. and Indiveri, G., Embedded neuromorphic vision for humanoid robots, Seventh IEEE Workshop on Embedded Computer Vision, 2011 Benosman, R., Ieng, S., Clercq, C., Bartolozzi, C., Asynchronous Frameless Event-based Optical Flow, IEEE Transactions on Neural Networks (submitted) Space variant event-driven sensor prototype Log Polar Mapping temporal contrast pixel illumination spatial derivative noise reduction fovea and periphery Temporal derivative, stability analisys of feedback loop (left) and measured traces corresponding to different biases (right) Adaptive photoreceptor and spatial derivative C. Bartolozzi1, F. Rea1, C. Clercq1,5, N. Mandloi1, G. Indiveri2, D.B. Fasnacht2, M. Hofstaetter3, G. Metta4, R. Benosman5 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045 0 1 2 3 Pulse(V) Pxl 2, Slow Pxl 1, Fast 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045 1 2 3 Time(s) Pulse(V) Pxl 1, Slow Pxl 2, Fast Simulation of velocity circuits Response to an edge traveling across two pixels at different speeds. The velocity is calculated as time-to-travel across neighbouring pixels Bias Generator Asynchronous Logic Asynchronous Logic Integrate & Fire Change Detector Light Intensity Meta pixel periphery fovea CD CD CD CD CD CD CD CD LI LI LI LI IF CD Dedicated embedded AER hardware Goal: to develop dedicated embedded infrastructure supporting efficient processing of asynchronous data GAEP SPARC-compatible general purpose Address-Event processor with 20-bit 10ns-resolution asynchronous sensor data interface in 0.18μm CMOS FPGA Orange: Interface logic, blue/striped: FIFOs, green: internal logic & control. Arrows: AER data-path (solid), bias control (dashed). iHead Miniaturized embedded board for AER processing, configuration and routing. Comprises the FPGA and the GAEP, CCSAER, USB and ETH connections DVS_L DVS_R FPGA DVS_L_AER DVS_nReset DVS_L_BG BG_PwrDown GAEP JTAG ETH USB BG commands AER data GAEP config & program FPGA code & debug iHEAD CCSAER USB-FX2 Flash Memory SRAM Clock Reset GAEP config DVS_R_AER DVS_R_BG AER data UART BG commands AER data Chip block diagram PC104 i7 CPU Router WiFi Optical flow Event-driven sensors transmit information about local changes in their field of view at the time they occur. The optical flow is updated for each event resulting in extremely sparse computation over time and space. Events from a bouncing ball Consecutive “frames” of accumulated events. Events at t i+1, events computed from optical flow at t i , lag between prediction and real position. Estimated optical flow The computation is frame-less and event-driven. Optical flow snapshot for a 5ms sequence Space-time representation of a rotating white disk with a black bar. Bar orientation estimated by the events-based (+) and by the frame-based (-) optical flow algorithms versus the true bar orientation over time. 1 Istituto Italiano di Tecnologia (IIT) - 2 Institute of Neuroinformatics (INI|UZH|ETH) 3 Austrian Institute of Technology (AIT) - 4 Università di Genova (UNIGE) 5 Institut de le Vision - Université Pierre et Marie Curie

Transcript of eMorph: towards neuromorphic robotic vision · A PCI based high-fanout AER mapper with 2 GiB RAM...

Page 1: eMorph: towards neuromorphic robotic vision · A PCI based high-fanout AER mapper with 2 GiB RAM look-up table, 0.8 us latency and 66MHz output event-rate, Conference on Information

eMorph: towards neuromorphic robotic vision

Information technology has not yet delivered arti�cial

systems that can compare with biology in reliably, robustly and e�ciently

extracting information from the often noisy and ambigu-ous real world, and interact-

ing with the world by generating appropriate

behaviours.

Tracker Motion Sensor- measures contrast, spatial and temporal derivative and velocity of moving edges- comprises circuits that implement a model of “selective attention” for a smart readout

eMorph is an EU funded project - ICT-FET 231467

Visual System Visual Primitives Technology

eMorph will deliver a methodology for event-driven arti�cial vision on a robotic demonstrator, giving the chance to the robotic

community to have a ready-to-use platform to directly interact and experience the power of neuromorphic tools.

Visual system integrationWe developed a series of printed circuit boards to integrate on the iCub event driven sensors and dedicated processors.

iCub navigationWe developed an autonomous vehicle with 6 holonomic, omnidirectional wheels.

Asynchronous, event-driven, space variant visual sensorsGoal: to design asynchronous, data-driven, biologically inspired, vision sensors with non-uniform morphology, using analog VLSI neuromorphic circuits

High level event-driven visual algorithmsGoal: to develop a supporting data-driven asynchronous computational paradigm for machine-vision that is radically di�erent from conventional image processing

System integrationGoal: to validate the system on the humanoid robot iCub �tted with the neuromorphic sensors, and controlled by the asynchronous machine-vision algorithms in real-time behavioral tasks

PublicationsBartolozzi, C. and Indiveri, G., Selective Attention in Multi-Chip Address-Event Systems, Sensors, 2009Orchard, G., Bartolozzi, C. and Indiveri, G., Applying neuromorphic vision sensors to planetary landing tasks, IEEE Biomedical Circuits and Systems Conference, 2009. Hofstätter, M., Schön, P. and Posch, C., A SPARC-compatible general purpose Address-Event processor with 20-bit 10ns-resolution asynchronous sensor data interface in 0.18μm CMOS, IEEE International Symposium on Circuits and Systems, 2010.Posch, C., Matolin, D., Wohlgenannt, R., Hofstaetter, M., Schoen, P., Litzenberger, M., Bauer, D., and Garn, H., Live Demonstration: Asynchronous Time-Based Image Sensor (ATIS) Camera with Full-Custom AE Processor, IEEE International Symposium on Circuits and Systems, 2010. Awarded as best live demonstrationFasnacht, D.B. and Indiveri, G., A PCI based high-fanout AER mapper with 2 GiB RAM look-up table, 0.8 us latency and 66MHz output event-rate, Conference on Information Sciences and Systems (CISS 2011), 2011.Bartolozzi, C., Metta, G., Hofstaetter, M. and Indiveri, G., Event-driven vision for the iCub, Bio-Mimetic and Hybrid Approach to Robotics, Workshop at IEEE International Conference on Robotics and Automation, 2011.Bartolozzi, C., Mandloi, N.K. and Indiveri, G., Attentive motion sensor for mobile robotic applications, IEEE International Symposium on Circuits and Systems, 2011.Bartolozzi, C., Fasnacht, D.B., Rea, F., Hofstaetter, M., Clercq, C., Metta, G. and Indiveri, G., Embedded neuromorphic vision for humanoid robots, Seventh IEEE Workshop on Embedded Computer Vision, 2011Benosman, R., Ieng, S., Clercq, C., Bartolozzi, C., Asynchronous Frameless Event-based Optical Flow, IEEE Transactions on Neural Networks (submitted)

Space variant event-driven sensor prototypeLog Polar Mapping

temporal contrastpixel illumination spatial derivativenoise reductionfovea and periphery

Temporal derivative, stability analisys of feedback loop (left) and measured traces corresponding to di�erent biases (right)

Adaptive photoreceptorand spatial derivative

C. Bartolozzi1, F. Rea1, C. Clercq1,5, N. Mandloi1, G. Indiveri2, D.B. Fasnacht2, M. Hofstaetter3, G. Metta4, R. Benosman5

0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.0450

1

2

3

Pul

se(V

)

Pxl 2, SlowPxl 1, Fast

0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045

1

2

3

Time(s)

Pul

se(V

)

Pxl 1, SlowPxl 2, Fast

Simulation of velocity circuits Response to an edge traveling

across two pixels at di�erent speeds. The velocity is

calculated as time-to-travel across neighbouring pixels

Bias Generator

Asynchronous Logic

Asy

nchr

onou

s Lo

gic

Integrate & Fire

Change Detector

Light Intensity

Meta pixel

periphery

fovea

CD CD

CD CDCD

CD CDCD

LILI

LI

LI

IF

CD

Dedicated embedded AER hardwareGoal: to develop dedicated embedded infrastructure supporting e�cient processing of asynchronous data

GAEP SPARC-compatible general purpose Address-Event processor

with 20-bit 10ns-resolution asynchronous sensor data interface in 0.18µm CMOS

FPGA Orange: Interface logic, blue/striped: FIFOs, green: internal logic & control. Arrows: AER data-path (solid), bias control (dashed).

iHeadMiniaturized

embedded board for AER

processing, con�guration

and routing. Comprises the FPGA and the GAEP, CCSAER, USB and ETH connections

DVS_L

DVS_RFPGA

DVS_L_AER

DVS_nReset

DVS_L_BG

BG_PwrDown

GAEP

JTAG

ETH

USB BG commandsAER data

GAEP con�g& program

FPGA code& debug

iHEAD

CCSAER

USB-FX2

Flash Memory SRAM

Clock

Reset

GAEP con�g

DVS_R_AERDVS_R_BG

AER

dat

a

UART

BG commandsAER data

Chip block diagram

PC104

i7 CPU

RouterWiFi

Optical �owEvent-driven sensors transmit information about local changes in their �eld of view at the time they occur. The optical �ow is updated for each event resulting in extremely sparse computation over time and space.

−20 0 20 40 60 80 100 120 140−20

0

20

40

60

80

100

120

140

Events from a bouncing ballConsecutive “frames” of accumulated events.

Events at ti+1, events computed from optical �ow at ti, lag between prediction and real position.

Estimated optical �owThe computation is frame-less and

event-driven.

Optical �ow snapshot for a 5ms sequence

Space-time representation of a rotating white disk with a black bar. Bar orientation estimated by the events-based (+) and by the frame-based (-) optical �ow algorithms versus the true bar orientation over time.

1 Istituto Italiano di Tecnologia (IIT) - 2 Institute of Neuroinformatics (INI|UZH|ETH)3 Austrian Institute of Technology (AIT) - 4 Università di Genova (UNIGE)

5 Institut de le Vision - Université Pierre et Marie Curie