Low-power smart imagers for vision-enabled Low-power smart imagers for vision-enabled wireless sensor networks and a case study wireless sensor networks and a case study J. Fernández-Berni, R. Carmona-Galán, Á. Rodríguez-Vázquez Institute of Microelectronics of Seville (IMSE-CNM), CSIC - Universidad de Sevilla (Spain) First ACM/ I EEE I nternational Workshop On Architecture of Smart Camera Clermont Ferrand, France April 5-6, 2012
Low-level image processing Low-level image processing INSTRUCTION FLOW, COMPUTATIONAL LOAD Sobel Operators - Potential parallel operation - Moderate accuracy required WASC April 2012, Clermont-Ferrand (France)
Conventional approach Conventional approach Analytic issues are mostly software issues Brute force pattern matching used by many system developers Extremely inefficient in terms of speed and power Imager Digital Signal Processor Low-level Mid-level High-level ADC tasks tasks tasks F F f f I nformation Flow: F F >> >> f > f > f ARRAY OF SENSORS WASC April 2012, Clermont-Ferrand (France)
Focal-plane array computing Focal-plane array computing Content-aware sensing-processing Progressive extraction of relevant information Parallel and distributed processing Distributed memory Smart Imager Digital Signal Processor Mid-level High-level ADC tasks tasks f f f I nformation Flow: f > f > f ARRAY OF SENSOR – PROCESSORS WASC April 2012, Clermont-Ferrand (France)
Focal-plane array computing Focal-plane array computing Sensor Memory P ER P IXEL : Optical sensing Neighbor Connectivity Local Processing Mixed ‐ signal processor Local Memory Single Instruction Multiple Data (SIMD) Architecture WASC April 2012, Clermont-Ferrand (France)
Focal-plane array computing Focal-plane array computing Several generation of chips designed FLIP-Q With fully programmable features Covering a large variety of functional targets Image-to-Decision at >1,000F/s with 60nW per pixel Spatio-temporal filtering with 22nJ/cycle Content-aware HDR acquisition with >145dB intra-frame DR Etc. WASC April 2012, Clermont-Ferrand (France)
J. Fernández Berni, R. Carmona Galán and L. FLI P-Q: floorplan FLI P-Q: floorplan Carranza González, “FLIP-Q: A QCIF Resolution Focal-Plane Array for Low-Power Image Processing,” in IEEE J. Solid-State Circuits, vol. 46, no. 3, pp. 669–680, March 2011 WASC April 2012, Clermont-Ferrand (France)
FLI P-Q: elementary processing cell FLI P-Q: elementary processing cell • Reset transistor • n-well/p-substrate photodiode • Electronic global shutter • Programmable block-wise image filtering and averaging • Programmable block-wise image energy computation • Readout circuitry WASC April 2012, Clermont-Ferrand (France)
Physical design Physical design Crucial aspect affecting the total area, the fill factor and the pixel pitch The electrical design must be realized bearing in mind the 29µm subsequent physical design Relevant issues: Metal layers available Full-custom routing Make the most of the design rules 34µm WASC April 2012, Clermont-Ferrand (France)
Physical design Physical design 60% +2% per additional µm in the elementary cell WASC April 2012, Clermont-Ferrand (France)
FLI P-Q: A prototype smart imager FLI P-Q: A prototype smart imager WASC April 2012, Clermont-Ferrand (France)
FLI P-Q: on-chip early vision FLI P-Q: on-chip early vision Programmable Gaussian filtering chip ideal error WASC April 2012, Clermont-Ferrand (France)
FLI P-Q: on-chip early vision FLI P-Q: on-chip early vision Fully-programmable multi-resolution scene representation On-chip images WASC April 2012, Clermont-Ferrand (France)
FLI P-Q: on-chip early vision FLI P-Q: on-chip early vision Image pre-distortion for reduced kernel filtering Original Reduced kernels kernels WASC April 2012, Clermont-Ferrand (France)
Wi-FLI P: a vision-enabled WSN node Wi-FLI P: a vision-enabled WSN node Smart Imager Digital Signal Processor Mid-level High-level ADC tasks tasks f f f I nformation Flow: f > f > f ARRAY OF SENSOR – PROCESSORS WASC April 2012, Clermont-Ferrand (France)
Wi-FLI P: a Wi-FLI P: a vision-enabled vision-enabled WSN node WSN node Imote2 (MEMSIC Inc.) WASC April 2012, Clermont-Ferrand (France)
Wi-FLI P: a vision-enabled WSN node Wi-FLI P: a vision-enabled WSN node WASC April 2012, Clermont-Ferrand (France)
Wi-FLI P: a vision-enabled WSN node Wi-FLI P: a vision-enabled WSN node WASC April 2012, Clermont-Ferrand (France)
Wi-FLI P: a vision-enabled WSN node Wi-FLI P: a vision-enabled WSN node DoG-based edge detection WASC April 2012, Clermont-Ferrand (France)
Wi-FLI P: a vision-enabled WSN node Wi-FLI P: a vision-enabled WSN node Very low throughput due to slow GPIO ports and TinyOS latency WASC April 2012, Clermont-Ferrand (France)
Case study: early detection of forest fires Case study: early detection of forest fires High economic cost Short maintenance cycles Coarse grain coverage Exact location must be inferred 27
Case study: early detection of forest fires Case study: early detection of forest fires • Vision-enabled Wireless Sensor Network ADVANTAGES DRAWBACKS Robustness Ultra low power consumption required Scalability Reliability Better temporal resolution Simpler smoke location 28
Case study: early detection of forest fires Case study: early detection of forest fires • Reconfigurable focal plane • A power-efficient vision algorithm for smoke detection • Multiresolution scene representation Candidate regions • Clustering ratio • Growth rate • Propagation speed SMOKE! 29
Case study: early detection of forest fires Case study: early detection of forest fires • Preliminary field tests http://www.imse-cnm.csic.es/vmote/ Original sequence Motion detector Our algorithm 30
Case study: early detection of forest fires Case study: early detection of forest fires • On-site smoke detection with Eye-RI S v1.2 31
Case study: early detection of forest fires Case study: early detection of forest fires • Field tests with Wi-FLI P 32
Case study: early detection of forest fires Case study: early detection of forest fires • Prescribed burning of a 95m x 20m shrub plot Wi-FLIP monitored all the activity for over two hours No false alarm triggered Successful smoke detection for two of the three vegetation areas explored Thin smoke generated from a very sparse vegetation area was not detected 33
CONCLUSI ONS CONCLUSI ONS • Early vision tasks represent a considerably heavy computational load . • SIMD-based massively parallel mixed-signal processing takes advantage of their intrinsic characteristics to achieve high power efficiency and computational power . • FLI P-Q : A prototype vision chip tailored for ultra low-power applications. Very competitive in the state of the art. • Wi-FLI P : A vision-enabled Wireless Sensor Network node supported by I mote2 . Current drawback: low throughput. • Case study: Early detection of forest fires , with very good results in terms of reliability . 34
Thank you very much Thank you very much for your attention for your attention berni@imse-cnm.csic.es berni@imse-cnm.csic.es Publication Date: May 31, 2012 35
Acknowledgments This work is financially supported by Andalusian Regional Government, through project 2006-TIC-2352, the Spanish Ministry of Economy and Competitiveness, through projects TEC 2009-11812 and IPT-2011-1625-430000, both co-funded by the EU-ERDF and by the Office of Naval Research (USA), through grant N000141110312. 36
Recommend
More recommend