sensor-level privacy for thermal cameras - focus...
TRANSCRIPT
Sensor-level Privacy for Thermal Cameras
Francesco Pittaluga
Aleksandar Zivkovic
Sanjeev J. Koppal
University of Florida
Imaging and Tracking People
Surveillance Military
Gaming IoT Mobile
2Intro Digitization Noise Exposure Conclusion
Balancing Privacy and Utility
Hospitals Schools
Retirement homes Workplaces
3Intro Digitization Noise Exposure Conclusion
Conventional Privacy Processing
Scene
Camera
ComputerStorageSecure
Data
4
Post-capture processing Lock data with cryptography
Edit images computationally
Boyle 2000, Sweeney 2002, Johnson et al. (IETF) 2003,
Gross et al. 2009, Agrawal and Narayanan 2011, …
Intro Digitization Noise Exposure Conclusion
Conventional Privacy Processing
Scene
Camera
ComputerStorageSecure
Data
Post-capture strategy has an inherent vulnerability
5Intro Digitization Noise Exposure Conclusion
Privacy Preserving Computational Cameras
Scene
Only capture light-field samples that are needed
Computational Camera
Secure
Data
6Intro Digitization Noise Exposure Conclusion
Computational Camera: Three Approaches
SceneComputational Camera
Secure
DataComputerStorage
X Pattern recognition
in every frameX Special
OpticsX New sensor
7
Chattopadhyay and Boult 2007, Winkler and Renner 2010, Narayanan and Mrityunjay 2011
Nelson et al. 2005,Winkler et al. 2014, Fernandez-Berni et al. 2014
Three Problems Limited Adoption
Zhang et al, 2014, Pittaluga and Koppal 2015
Intro Digitization Noise Exposure Conclusion
Our Idea: Sensor-level Privacy for Thermal Cameras
Scene
Our idea: use a thermal sensor
(Wavelengths > 3µm)
Computational Camera
Secure
DataComputerStorage
8
Reliable Physics
Based Processing Off the Shelf
Optics
Thermal
sensor
Intro Digitization Noise Exposure Conclusion
Thermal Cameras are Coming Soon!
Thermal cameras are not exotic anymore
Thermal face recognition works
FLIR One 80x60
~$250
Melaxis 16x4
~$75FLIR A6751SC
9Intro Digitization Noise Exposure Conclusion
Three Sensor Level Approaches
SceneComputational Camera
Secure
Data
Thermal Sensor
10Intro Digitization Noise Exposure Conclusion
Three Sensor Level Approaches
Thermal Sensor
Digitization Noise Exposure
11Intro Digitization Noise Exposure Conclusion
Three Sensor Level Approaches
Thermal Sensor
Digitization Noise Exposure
12Intro Digitization Noise Exposure Conclusion
Humans are broadband
13
Spectral Power Response vs Wavelength
8
11
7 14
Human Spectral Response
r(λ)
(μm))
𝑊
𝑚2 𝑠𝑟 𝜇𝑚
Intro Digitization Noise Exposure Conclusion
Facial skin temperature reaches an equilibrium
37 Degrees
(Celsius)
Outside temperature known
= T Celsius
Facial skin temperature F (T, 37)
14Intro Digitization Noise Exposure Conclusion
This mapping is known
Olesen and Parsons 2002
15
Skin
Tem
pera
ture
°C
Ambient Temperature °C
Intro Digitization Noise Exposure Conclusion
Removing pixels in this range during digitization
ASIC Modification
16
Pixel
Readout
Output
Voltage
A-to-D
Masking measurements
based on temp. range
16-bit
decoder
Lower
Voltage
Upper
VoltageUpper Bound
Comparator
Lower Bound
Comparator
AND
AND
NOT
Seq. Counter
Random-Access
Memory
Address (18.0)
Data(15.0)
Intro Digitization Noise Exposure Conclusion
Digitization Result
17Intro Digitization Noise Exposure Conclusion
Digitization Result
18Intro Digitization Noise Exposure Conclusion
Three Sensor Level Approaches
Thermal Sensor
Digitization Noise Exposure
19Intro Digitization Noise Exposure Conclusion
Adding noise to the bolometer
20
GFID
GSK
VCC
Blind Bolometer
Active Bolometer
Vbus
Vout
Cint
Tunable Bias
Voltages
Intro Digitization Noise Exposure Conclusion
The effect of bias voltages
21
GSK
GFID
5 V
5 V
0 V
0 V
GSK
GFID
5 V
5 V
0 V
0 V
GSK
GFID
5 V
5 V
0 V
0 V
GSK
GFID
5 V
5 V
0 V
0 V
Exposure 5 Exposure 15
Exposure 25 Exposure 75
Intro Digitization Noise Exposure Conclusion
Calibrating for noise and privacy
22
GSK
GFID
5 V
5 V
0 V
0 V
9000
0
0 120Grayscale values
Histogram of values for the highest standard deviation
Occurrences
2𝜎 = 14 graylevels
Bias voltages with exposure set to 5
For a flat lambertian plane
Intro Digitization Noise Exposure Conclusion
Noise Result: Head Tracking
23Intro Digitization Noise Exposure Conclusion
Three Sensor Level Approaches
Thermal Sensor
Digitization Noise Exposure
24Intro Digitization Noise Exposure Conclusion
Temperature and radiant power
25
Spectral Power Response vs Wavelength
8
11
7 14
s(λ)
Camera sensitivity
(μm)
𝑊
𝑚2 𝑠𝑟 𝜇𝑚
Pixel Number vs Radiant Power
Φ𝑐𝑜𝑙𝑑 Φℎ𝑜𝑡
𝐼max~
𝐼min~
𝐼max
Φ = λ𝑡
λℎ
𝑠 λ 𝑟 λ 𝑑λ
Human Spectral Response
r(λ)
Φℎ𝑢𝑚𝑎𝑛
r(λ)r(λ)
Intro Digitization Noise Exposure Conclusion
No capture region
26
Spectral Power Response vs Wavelength
8
11
7 14
Φ𝑚𝑖𝑛 = λ𝑡
λℎ
𝑠 λ [𝑟 λ −∆ 𝜆
2] 𝑑λ
Φ𝑚𝑎𝑥 = λ𝑡
λℎ
𝑠 λ [𝑟 λ +∆ 𝜆
2] 𝑑λ
(μm)
𝑊
𝑚2 𝑠𝑟 𝜇𝑚
Pixel Number vs Radiant Power
Φ𝑚𝑖𝑛 Φ𝑚𝑎𝑥
𝐼max
“No
Ca
ptu
re”
s(λ)
Camera sensitivity
Δ(𝜆)Human Spectral Response
r(λ)
Intro Digitization Noise Exposure Conclusion
Exposures that remove no capture region
27
Spectral Power Response vs Wavelength
8
11
7 14
s(λ)
Camera sensitivity
(μm)
𝑊
𝑚2 𝑠𝑟 𝜇𝑚
Underexposure
Overexposure
Pixel Number vs Radiant Power
Φ𝑚𝑖𝑛 Φ𝑚𝑎𝑥
𝐼max~
𝐼min~
𝐼max
“No
Ca
ptu
re”
𝑡 ≥𝑔 𝐼𝑚𝑎𝑥(𝑔)
Φ𝑚𝑖𝑛
𝑡 ≤𝑔 𝐼𝑚𝑖𝑛(𝑔)
Φ𝑚𝑖𝑛
Human Overexposed Human Underexposed
Human Spectral Response
r(λ)
Intro Digitization Noise Exposure Conclusion
Optimal algorithm to obtain exposures
28
Grossberg and Nayar 2003
𝝽(𝑛, 𝑇) = Γ𝑚𝑖𝑛
Φ𝑚𝑖𝑛
ℎ𝑑𝑒𝑠′ − ℎ′ 𝑝𝜔𝑑Φ +
Φ𝑚𝑎𝑥
Γ𝑚𝑎𝑥
ℎ𝑑𝑒𝑠′ − ℎ′ 𝑝𝜔𝑑Φ 𝜔 =
0 ℎ𝑑𝑒𝑠′ (Φ) < ℎ′(Φ)
1 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒,
…
Error Function Binary Weights
1. 𝑇𝑖 > 0
2. ∀ 𝑇𝑖 𝑇𝑖 ≥ 𝐼𝑚𝑎𝑥
Φ𝑚𝑖𝑛⊕ 𝑇𝑖 ≤
𝐼𝑚𝑖𝑛
Φ𝑚𝑎𝑥
𝑎𝑟𝑔𝑚𝑖𝑛 𝑛,𝑇 𝝽(𝑛, 𝑇) 𝑠. 𝑡.
Intro Digitization Noise Exposure Conclusion
Optimal algorithm to obtain exposures
Algorithm assumes a single no capture region
Brute force search is therefore tractable
29
Objective function score
Optimal solution for
112 exposures
Grid search index
HDR Image
Intro Digitization Noise Exposure Conclusion
HDR Results
Over Under Fusion
30Intro Digitization Noise Exposure Conclusion
HDR Results
Over Under Fusion
31Intro Digitization Noise Exposure Conclusion
HDR Results
Over Under Fusion
32Intro Digitization Noise Exposure Conclusion
HDR Results
33Intro Digitization Noise Exposure Conclusion
Method Comparison
34Intro Digitization Noise Exposure Conclusion
Comparison
Digitization Noise Exposure
• Low noise
• Good image
quality
• Real-time
• Hardware and
firmware upgrades
• Low noise
• No hardware
modification
• Good image quality
• Multiple Images and
more capture time
• Real-time
• No hardware
modification
• Low image quality
• Noisy
35Intro Digitization Noise Exposure Conclusion
Future Work
Pilot deployment program of private sensors at
UF Health Shands Hospital.
Generate database of private face images to for
privacy challenge.
Generate database of private videos for activity
recognition in a hospital setting.
36Intro Digitization Noise Exposure Conclusion
DHS
Sanjeev Koppal Andreas Enqvist
This material is based upon work supported by theU.S. Department of Homeland Security under Grant Award
Number, 2014-DN-077-ARI083-01. The views and conclusionscontained in this document are those of the authors
and should not be interpreted as necessarily representing theofficial policies, either expressed or implied, of the U.S. Department
of Homeland Security.
Acknowledgements
Intro Digitization Noise Exposure Conclusion
Summary: Three Sensor Level Approaches
Digitization Noise Exposure
38Intro Digitization Noise Exposure Conclusion