Industrial DVS Design; Key Features and...
Transcript of Industrial DVS Design; Key Features and...
Industrial DVS Design;Key Features and Applications
Hyunsurk Eric Ryu, PhD
System LSI, Samsung Electronics
Hyunsurk Eric Ryu Industrial DVS Design 1 of 27
Outline
Industrial DVS DesignHyunsurk Eric Ryu
Samsung S.LSI Dynamic Vision Sensor
Key Features
◼ Low Latency
◼ Minimized Motion Artifacts
◼ Anti-Flicker
◼ Low Power Operation
Applications
◼ Sparse Edge-based Object Recognition
◼ Pose Estimation with Visual Information: DVS-*SLAM
Summary and Discussion*SLAM: Simultaneous Localization and Mapping
2 of 27
DVS Gen1 (R&D Ver., 2014)
Industrial DVS DesignHyunsurk Eric Ryu
Brief Specifications
◼ 640 x 480 Pixel Array with 1/2.5-inch optics
◼ Pixel size : 9 um
◼ Dynamic range : 66 dB (5~10,000 Lux)
◼ Max. event processing rate : 6.5 Meps*
◼ Minimum detectable contrast (50% response)
: < 19%
◼ Interface : 20-bit Parallel
◼ Typical power consumption : 15 mW
Key Features
◼ Small pixel size (9 um)
9.7mm
8.0
mm
640 x 480 Pixel Array
Column Arbiter + Addr. Decoder
Ro
w A
rbite
r + A
dd
r. De
co
de
r
Co
ntro
lL
og
ic +
AM
UX
Bia
s
Ge
n.
Pixel
PD
9um
9u
m
*eps: event per second
Full custom logic design
Fabricated using the Samsung 90-nm Back Side Illumination (BSI) CIS process
3 of 27
DVS Gen2 (R&D Ver., 2016)
Industrial DVS DesignHyunsurk Eric Ryu
(ISSCC ’17)
Brief Specifications
◼ 640 x 480 Pixel Array with 1/2.5-inch optics
◼ Pixel size : 9 um
◼ Dynamic range : 90 dB (3~100,000 Lux)
◼ Max. event processing rate : 300 Meps
◼ Minimum detectable contrast (50% response)
: < 19%
◼ Interface : MIPI(1Gbps 4-lane), USB connectable
Parallel, I2C
◼ Typical power consumption : 80 mW
Key Features
◼ Standard digital I/F IP supported
◼ Digitally synthesized G-AER* for high throughput*G-AER: Group Address Event Representation
High data throughput using digitally synthesized G-AER scheme
4 of 27
DVS Gen3 (Product Ver., 2018)
Industrial DVS DesignHyunsurk Eric Ryu
Motion image artifact minimized by using global hold and global reset
Timestamp error minimized by applying sequential column scan readout
8.0mm
5.8
mm
640 x 480 Pixel Array
Bias Generator
Pixel Layout
PD
Digital Scan Logic
Event Storage
9um
9u
m
Dig
ital L
og
ic &
Fu
nctio
n B
locks
Brief Specifications
◼ 640 x 480 Pixel Array with 1/2.5-inch optics
◼ Pixel size : 9 um
◼ Dynamic range : 90 dB (3~100,000 Lux)
◼ Effective frame rate : > 2,000 fps (MIPI)
◼ Minimum detectable contrast (99.9% response)
: < 27.5%
◼ Interface : MIPI(1Gbps 4-lane), USB connectable
Parallel, I2C
◼ Typical power consumption : 65 mW
Key Features
◼ Global hold, Global reset, Column scan readout
5 of 27
DVS Gen4 (R&D Ver.)
Industrial DVS DesignHyunsurk Eric Ryu
Brief Specifications
◼ 1280 x 960 Pixel Array with 1/2-inch optics
◼ Pixel size : 4.95 um
◼ Dynamic range : 90 dB (3~100,000 Lux)
◼ Frame rate : fixed 1,000 fps (MIPI)
◼ Minimum detectable contrast (99.9% response)
: < 27.4%
◼ Typical power consumption : 140 mW
Key Features
◼ Pixel circuit split using two-stack wafer bonding
◼ 2nd CCI supported, CIS RGB Format, De-noise,
Anti-Flicker
8372.7um
76
40
.1u
m
1280 x 960 Pixel Array (TOP/BOT)
4.95um
Digital Scan / Core Logic (BOT)
AP
S-D
igit
al
Inte
rfa
ce
(T
OP
)
AP
S-D
igit
al
Inte
rfa
ce
(T
OP
)
Bias GenTIA
ADC
MIPIOTP ABBG TEMP
Event Storage
GH
LD
Comparator Output Logic + Reset Logic
ON
Output
Logic
REQY
ACKY
Reset
LogicRST
OFF
ON_HLD
OFF_HLD
SE
LX
M
Capacitive-feedback
Amplifier (CapAMP)
-A2
C1 C2
Source
Follower (SF)PD + Log I-V Amp (LogAMP)
MFBN
MSF
VLOG
IPD
ISF
-A1VPD
VSF VOUT
GR
ST
TOP (2.8V)
→ BOT (1.8V)
MFBPVBIAS
(simplified) Pixel Diagram
상판 pixel (4x4) C2C (4x4) 하판 pixel (4x4)
Event Storage
GH
LD
Comparator Output Logic + Reset Logic
ON
Output
Logic
REQY
ACKY
Reset
LogicRST
OFF
ON_HLD
OFF_HLD
SE
LX
M
Capacitive-feedback
Amplifier (CapAMP)
-A2
C1 C2
Source
Follower (SF)PD + Log I-V Amp (LogAMP)
MFBN
MSF
VLOG
IPD
ISF
-A1VPD
VSF VOUT
GR
ST
TOP (2.8V)
→ BOT (1.8V)
MFBPVBIAS
(simplified) Pixel Diagram
상판 pixel (4x4) C2C (4x4) 하판 pixel (4x4)
4.9
5u
m
2x2 Pixel
(Bottom)
2x2 Pixel
(Top)
─ Top─ Bot─ Top/Bot
Pixel circuit split by using two-stack wafer bonding for smaller pixel size
Event signal processing block (Anti-flicker, De-noise , etc.) added
6 of 27
DVS Gen2 Results, only valid for small events
Industrial DVS Design 7 of 27Hyunsurk Eric Ryu
DVS Scene Characteristics
Hyunsurk Eric Ryu Industrial DVS Design 8 of 27
0
200
400
600
800
1000
1200
1400
1600
1800
0 10 20 30 40 50 60 70 80
Time (sec)
Eve
nt Rate
(ke
ps)
Event Probability
Hyunsurk Eric Ryu Industrial DVS Design 9 of 27
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
0 10 20 30 40 50 60 70
Even
t P
rob
abili
ty
Temporal Contrast (%)
100Lux
5 Lux
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
0 5 10 15 20 25 30
Even
t P
rob
abili
ty
Temporal Contrast (%)
…Pixel Array
…
unfair arbitration
Column selection
Column Arbiter
Artifacts and delay by unfair arbitration
Industrial DVS Design 10 of 27Hyunsurk Eric Ryu
Event scanning with a capacitor event memory
…Pixel Array
Fast column sequential read
Industrial DVS Design 11 of 27Hyunsurk Eric Ryu
Industrial DVS DesignHyunsurk Eric Ryu
Original AER handles the individual pixel data with address, polarity, and event generation time.
Group addressing reduces the latency by the interface bandwidth limitation
Pixel Array(640 x 480)
DVS with Event-Group Handshaking Readout
Sam
pling C
ircuit
Dig
itally S
ynth
esiz
ed
Gro
up *
AER
E[0]
E[1]
E[479]
Column Arbiter
CReq
AReq
CAddr[9:0] CAddr_BUF[9:0]
GAddr[5:0]
GData[7:0]
COL_yyyCOL_xxx
CAddr [9:0]
CAddr_BUF[9:0]
COL_yyyCOL_xxx
CReq
AReq
GAddr[5:0]
CLK
GData[7:0]
AER induced latency
*AER: Address Event Representation
12 of 27
AER induced motion artifact
Industrial DVS DesignHyunsurk Eric Ryu
Image artifact could be induced by the mismatch between event generation
time and readout time under high event rate condition
320 320
160 160 160 160
640
…Pixel Array
480
…
Arbitration
Arbitrary column selection
Column Arbiter
640
…Pixel Array
480
…
Sequential Scanning
Sequential column selection
Column Scanner
1
10
100
1000
1 10 100 1000
Sequential Column Selection
Word-serial AER using Arbiter
Tim
esta
mp Err
or
(Late
ncy)
(μsec)
Average Event Rate (Mega event per second)
Input: Poisson random events
13 of 27
Global Hold, Global Reset
Industrial DVS DesignHyunsurk Eric Ryu
Global Hold Off Global Hold On
rotating fan@8,000rpm
rotating fan@8,000rpm
Global Reset Off Global Reset On
One global reset every 32 frames
Tail Events Suppressed Tail Events
Jello Effect
Global hold is implemented with an event storage in each pixel and its global
control signal
To minimize unwanted tail event, global pixel-reset function is applied
14 of 27
Imbedded Anti-Flicker
Industrial DVS DesignHyunsurk Eric Ryu
Event Pass Rate vs. Flicker Frequency
Event Pass Rate vs. Non-Flicker Recordings
Average Non-flicker Event Pass Percentage
97.2%
Shaking Hand before Flickering Monitor
Walking person(AF @non-flickering environment)
15 of 27
Low Power Operation
Industrial DVS DesignHyunsurk Eric Ryu
Activity Detection Spatial Histogram
32
32
(0,0) (0,1) (0,2) (0,3) (0,16) (0,17) (0,18) (0,19)
(1,0) (1,1) (1,2) (1,3) (1,16) (1,17) (1,18) (1,19)
(14,0) (14,1) (14,2) (14,3) (14,16) (14,17) (14,18) (14,19)
One patch comprises 32x32 pixel array
16 of 27
DVS Human Detection
Industrial DVS DesignHyunsurk Eric Ryu
Recognition with securing privacy
Scene DVS RecognitionDVS Output
(Edge Image)
17 of 27
Industrial DVS DesignHyunsurk Eric Ryu
(person, x:230, y:410) (dog, x:70, y:432) (cat, x:321, y:125)
(chair, x:322, y:268) (box, x:120, y:122) (electric fan, x:621, y:325)
Total 850K images (11 categories) are used for network training.
1M-labeled VGA data set, 2.5x faster training and easy converging
Smaller Database and Faster Training
18 of 27
Industrial DVS DesignHyunsurk Eric Ryu
Algorithm Flow
Input Event Stream
Feature Extractor
(Densenet,Darknet)
RegionProposalBranch
Refine-ROIBranch
FRCNN structure
Event Pre-process
HumanDetection
Spatial and Temporal
voting
CountingResult
C++ based Caffe v1 with CPU only mode (no CUDA lib)
◼ DVS event pre-process
Use historic event timestamp of DVS sensor to compensate the sparsity issue.
◼ Deep learning Based Human detection
Faster Region Convolutional Neural Network (FRCNN)
Calculate the position of objects and classify objects.
◼ Spatial and temporal voting
Fusion different detection results to improve accuracy.
19 of 27
Industrial DVS DesignHyunsurk Eric Ryu
Distance ≥ 10 lux 5 ~ 10 lux
Recall
0.5 ~ 1.0m 98% 96%
1.0 ~ 5.0m 99% 98%
5.0 ~ 7.0m 96% 92%
False Alarm Rate (FAR) 1%
Accuracy Test FAR Test
Human Detection in Edge Device (ARMv8 Board)
92ms @ Exynos 7570, 2 cores (60%)
20 of 27
Smaller Network with DVS Images
Industrial DVS DesignHyunsurk Eric Ryu
Sparsity and binary features of DVS images
◼ Few kernel numbers
◼ Few convolutional layers
◼ Large stride for layers in the front
Small and fast network without accuracy degradation.
DVS-Based human detection solution (smart home/IoT)
◼ Computational speed: 140ms/frame @Exynos 7570(1.4GHz, 8% CPU usages)
FRCNN+FPN# of
*Layers*FLOPs
ProcessingTime (ms)
CIS-based 91 5.8G 172
DVS-based 24 81M 15
➔~11.4x speed up
*backbone network**computed on Titan XCIS DVS
21 of 27
Industrial DVS DesignHyunsurk Eric Ryu
DVS SLAM
Durrant-Whyte, Bailey: “Simultaneous Localisation and Mapping ( SLAM ) : Part I The Essential Algorithms”
VO/VIO/VI-SLAM Algorithms
◼ Track location of device in unknown environment
◼ Use visual and inertial sensors
Challenges with standard image sensors
◼ Lack robustness to fast motion (blur)
◼ Lack robustness to HDR (loss of features)
◼ High latency (VR & AR, fast control loops)
DVS intrinsic properties deal with these
challenges naturally
D. Scaramuzza, “Visual Odometry and SLAM: past, present, and the robust-perception age”
22 of 27
DVS-SLAM Algorithm
Hyunsurk Eric Ryu Industrial DVS Design 23 of 27
Yes
DVS Events
IMU
DVS Framer IMU Integration
Frame
Pose EstimationKeyFrameSelection
1. FAST Feature Extraction2. 3D Map Points Projection to
Frame3. Feature Matching4. 3D-2D PnP (Perspective-n-Points)
Map Points Insertion
Using Triangulation
KeyFrame,Map Points
KeyFrames,Map Points
Bundle Adjustment
Optimizeddata
Failure Recovery
Map Points
to Pose Estimation
Location Manager
100 Hz 6-DoF Pose
Mid-termFrame based6DoF Pose
DVS-SLAM Video
Hyunsurk Eric Ryu Industrial DVS Design 24 of 27
DVS-SLAM provides robust tracking even at high-speed motions, thanks to high frame rate (>1000 FPS), sophisticated processing, low
computing load with DVS signal characteristics.
Industrial DVS DesignHyunsurk Eric Ryu
Measured Performances
Comparison with a CIS Platform (May 2019)
DVS-SLAM A CIS Solution
Average Median Average Median
Last Position Error (cm) 3.2 2.7 130 125
AAPE (cm) 2.7 2.3 45.4 39.8
Relative Drift Error (%) 0.28% 0.24% 13.61% 12.6%
Relative AAPE (%) 0.25% 0.16% 4.96% 4.02%
A CIS loses tracking
25 of 27
Industrial DVS DesignHyunsurk Eric Ryu
DVS+CIS SLAM
DVS + CIS: best of both worlds?
◼ DVS provides low latency visual sensing, robustness to fast
motion & HDR
◼ CIS provides precise long term feature tracking, loop closure
& re-localization ability
Sample performance of hybrid DVS+CIS SLAM on
“The Event-Camera Dataset”:
ScenarioVINS CIS Only
AAPE [m]VINS CIS + DVS
AAPE [m]
boxes_6dof 0.29 0.26
boxes_translation 0.09 0.07
hdr_boxes 0.18 0.16
poster_6dof 0.29 0.19
poster_translation 0.02 0.03
dynamic_translation 0.04 0.02
shapes_6dof 0.62 0.49
26 of 27
Summary and discussion
Industrial DVS DesignHyunsurk Eric Ryu
• There were several design changes for the latency, motion artifact, low
power operation, smaller pixel size operation, etc.
• DVS based visual recognition shows fewer and faster NN without loss of
classification accuracy.
• DVS SLAM shows robustness when the motion is fast, and/or for the high
dynamic range scene. DVS with CIS provides better performance than DVS
or CIS only cases.
• Remaining issues and research topics
• bandwidth minimization for higher event rate and spatial resolution
• improving the pixel sensitivity (response time) at dark illumination condition
• A way of high level visual information processing combined with SLAM
information
27 of 27