amr suleiman, zhengdong zhang, luca carlone, sertac...
TRANSCRIPT
Navion:AFullyIntegratedEnergy-EfficientVisual-Iner:alOdometryAcceleratorforAutonomousNaviga:onofNanoDronesAmrSuleiman,ZhengdongZhang,LucaCarlone,SertacKaraman,andVivienneSze
Demo:TrevorHendersonandDianaWo@
Session/Paper No. : 13.1
Mo:va:on
Virtualandaugmentedrealityonenergy-constraineddevices
SponsorsWanttoknowmore?Visitnavion.mit.edu
NavionResults
KeyChallenges• Large frame memories for feature
trackingandstereocomputaFon• Large graph memory with irregular
memoryaccess• Slowlinearsolver
OurApproach• Imagecompression
• 4.4xsmallerframememory• 4.9xlesspower
• Sparsefeaturetracksdatastructure• 5.4xsmallergraphmemory
• Sparselinearsolver• 5.2xsmallerlinearsolvermemory• 7.2xspeedup
Technology 65nmCMOSChiparea 4.0x5.0mm2
Logicgates 2,043kgatesSRAM 854KBFramesize 480x752Camera 28–171fpsKeyframe 16–90fpsPower 27– 43.2mW
PlaUorm Xeon[2]
ARM[2]
FPGA[2]
ASIC(Thiswork)
Accuracy* 0.20% 0.20% 0.24% 0.28%Cameramaxfps 21 5 20 171Keyframemaxfps 13 3 5 90Power(W) 26.10 2.33 1.46 0.024
NoGPS
Loca:on(Trajectory)
3DMap
UseimagesandinerFaldataforstatees0ma0on
5.0 mm
4.0
mm
Feature Tracking
Feature Detection
Undistort& Rectify
Undistort& Rectify
Sparse StereoLinear Solver
IFE
SharedMemory
Graph
Horizon States
VFE Control
Marginal
References[1]C.Forster,L.Carlone,F.Dellaert,D.Scaramuzza.“On-manifoldpreintegraFontheoryforfastandaccuratevisual-inerFalnavigaFon,”IEEETrans.RoboFcs,2016[2]Z.Zhang,A.Suleiman,L.Carlone,V.Sze,S.Karaman,“Visual-InerFalOdometryonChip:AnAlgorithm-and-HardwareCo-designApproach,”RSS2017.
VisionFrontend(VFE) Backend(BE)FeatureDetec:on(Iden:fyfeatures)• ExtractHarriscorners(upto200perframe)• Searchanddistributeacrossthewholeframe• Computeforeverykeyframe
FeatureTracking(Es:matemovementoffeatures)• Lucas-KanadepyramidalopFcalflow• Supportsupto3pyramidlevels• Computeforallframes(measuremovement
betweenkeyframes)
3DStereo(Es:matedepthoffeatures–a.k.a.landmarks)• Sparseblockmatching• Computeforeverykeyframe
FusesvisionandIMUesFmatestorefinethefinalstateesFmates(x’)
Mem
ory
size
3.5 MB 2.4 MB 1.6 MB 854 kB
1.5x 2.6x 4.1x
Image Compression
Two-stage Storage
Sparse LS Memory
Baseline
Frame Graph Linear Solver
Misc.
IMUFrontend(IFE)
AutonomousnavigaFonofnanodrones
x’1
x’2
x’3
Lk
Solvewithstate-of-the-artfactorgraphalgorithm
MinimizeinconsistenciesbetweenmeasurementsacrossFme
Updatestatees0mates(x’)
Over4000+factorsinop0miza0on
IMUFactors VisionFactorst=1
t=2 t=3
*MeasuredonEuRocMAVdataset
VIO:Visual-Iner:alOdometry
Gyroscope: EsFmateRotaFonAccelerometer: EsFmateTranslaFon
Pre-integra:on: IntegrateIMUmeasurementsintooneperkeyframe[1]
NavionChipArchitectureBackend Control
Data & Control BusBuild Graph
Linear Solver
Linearize
Marginal
Retract
GraphLinear Solver
Horizon StatesShared
Memory
Floating Point
Arithmetic
Matrix Operations
Cholesky
Back Substitute
Rodrigues Operations
Feature Tracking
(FT)
Previous FrameLine Buffers
Feature Detection
(FD)
Undistort & Rectify
(UR)
Undistort & Rectify
(UR)
Data & Control Bus
Sparse Stereo (SS)
Vision Frontend Control
RANSAC Fixed Point Arithmetic Point Cloud Pre-IntegrationFloating Point
ArithmeticIMU
memory
Current Frame
Left Frame
Right Frame
Vision Frontend (VFE)
IMU Frontend (IFE)
Backend (BE)
Register File
FullyIntegratedSystem
(nooff-chipprocessingorstorage)
VIOisaspecialinstanceofSimultaneousLocaliza6onandMapping(SLAM)