Progress in sensor performance testing, modeling and range prediction
using the TOD method: an overview
Piet Bijl, Maarten A. Hogervorst & Alexander Toet
TNO, Kampweg 5, 3769 DE, The Netherlands
ABSTRACT
The Triangle Orientation Discrimination (TOD) methodology includes i) a widely applicable, accurate end-to-end EO/IR
sensor test, ii) an image-based sensor system model and iii) a Target Acquisition (TA) range model. The method has
been extensively validated against TA field performance for a wide variety of well- and under-sampled imagers, systems
with advanced image processing techniques such as dynamic super resolution and local adaptive contrast enhancement,
and sensors showing smear or noise drift, for both static and dynamic test stimuli and as a function of target contrast.
Recently, significant progress has been made in various directions. Dedicated visual and NIR test charts for lab and field
testing are available and thermal test benches are on the market. Automated sensor testing using an objective synthetic
human observer is within reach. Both an analytical and an image-based TOD model have recently been developed and
are being implemented in the European Target Acquisition model ECOMOS and in the EOSTAR TDA. Further, the
methodology is being applied for design optimization of high-end security camera systems. Finally, results from a recent
perception study suggest that DRI ranges for real targets can be predicted by replacing the relevant distinctive target
features by TOD test patterns of the same characteristic size and contrast, enabling a new TA modeling approach. This
paper provides an overview.
Keywords: TOD, sensor test, sensor model, Target Acquisition, EO, IR, ECOMOS, EOSTAR
1. INTRODUCTION
Theoretical models and sensor tests that are able to supply Target Acquisition (TA) range performance with camera
systems are both essential elements in procurement processes and acceptance testing. TA models are used to
theoretically predict whether a certain camera system will meet the requirements. They calculate the expected TA
performance on the basis of the physical parameters of the sensor system. Sensor performance tests measure the
performance of a sensor system under well-defined (laboratory) circumstances and using well-defined stimuli, often with
a human-in-the-loop. They are required for (lab and field) acceptance of the delivered systems and their maintenance.
Ideally, a sensor performance model and a sensor performance test are two complementary parts of a single methodology
that are strongly connected and predict the same thing: field performance.
Current popular models and tests are:
• The Triangle Orientation Discrimination (TOD) model, based on the TOD test1.
• Thermal Range Model (TRM4), based on the Minimum Temperature Difference Perceived (MTDP)2 test
developed at FGAN-FOM (currently Fraunhofer) in Germany
• NV-IPM, based on the Targeting Task Performance (TTP) metric3,4, developed at NVESD in the US.
All three models are candidates to be included in an updated STANAG for nominal range performance prediction with
thermal imagers. The original STANAG 4347:19955 is cancelled. The update is currently under development by NATO
JCGISR Team of Experts Electro Optics and will include IR sensor modeling and testing. The TOD and MTDP tests are
candidates to be included as end-to-end thermal imager performance tests.
While the TTP approach is focused on modeling and also TRM4 contains a very detailed (analytical) sensor model, the
TOD was initially developed as an end-to-end sensor performance test1,6,7 combined with a range model8. A logical
choice of the (non-periodical) test pattern that represents target features, a robust bias-free forced-choice test and a
statistically sound analysis procedure form the basis of this very accurate and time-efficient test method. Ironically, these
differences with the standard EO/IR testing approach has hindered a wide acceptance and the development of test
equipment for a long time. Nevertheless, over the years the TOD test has been applied to a wide variety of image
forming systems including CCD cameras, thermal imagers, low light level cameras, image intensifiers, LADAR, X-ray
baggage screening systems, panoramic imagers, Helmet Mounted Displays and flight simulators. It has been applied to
static and dynamic conditions, to a variety of signal processing techniques, image compression, to sensors showing
smear and noise drift and even to characterize the effect of laser damage on performance9-34.
Validation studies of advanced use include: the effect of target contrast6,35, target motion, dynamic super resolution, local
adaptive contrast enhancement, sensor smear, combinations of these, and boost. The prediction error is always within
10% or within the experimental error of the validation data. In all TOD validation studies so far6,9,10,11,15,16,18,35, the
assumption that the TOD test is representative for TA performance has not been violated.
At the same time, several image-based TOD prediction models were developed: a receptor-based model, a neurologically
inspired model, and a simple correlation model36-38. Predictions were good but so far a standard model was never chosen.
A large advantage of an image-based approach over analytical modeling such as the TTP or TRM4 is that -in principle- it
allows performance characterization of non-linear systems with advanced image enhancement and or compression
techniques and may be extended to dynamic sensor performance and non-uniform scenes.
Recently, significant progress of the TOD methodology has been booked in all directions. First, interest in the test is
growing and the availability of test equipment is improving. Second, both an analytical and an image-based model have
been developed and implementation39-41 is under development. The models are based upon a systematically developed
TOD perception database. A key component of the image-based model is a Human Visual System (HVS) model that
mimics the human response in a TOD test task (i.e. judging the orientation of a set of triangle test pattern imaged by a
sensor under test). Finally, a recent experiment on the perception of ship targets35 has led to a new approach in TA range
modeling. Current TA models actually all make a range predictions for an ensemble of targets using set-dependent
criteria such as N50 (Johnson), V50 (TTP) and M75 (TOD model). Target sets are represented by a box of certain size and
contrast and without additional target shape or feature information. The new data suggest that ship detection and the
classification of ship features can be predicted by replacing those features by a TOD pattern of the same characteristic
size and (local) contrast. If this assumption is confirmed in validation studies, this may lead to DRI range predictions
based on individual target characteristics providing new opportunities for example in signature management.
This paper provides an overview of the TOD progress, bringing together results from recent papers. Section 2 addresses
the availability of EO/IR test equipment. In Section 3, the systematic approach of the development and validation of an
analytical and an image-based TOD model is described. Section 4 discusses the new approach in Target Acquisition
(TA) range modeling, and Section 5 shows recent developments in applications such as the European Target Acquisition
model ECOMOS40 and in the EOSTAR TDA41 and the use of image-based end-to-end modeling in system design. In
Section 6, a conversion from MTDP to TOD is deduced. Discussion and Conclusions are provided in Section 7.
2. TEST METHOD AND EQUIPMENT
The Triangle Orientation Discrimination (TOD)1,6,7 test was originally published in 1998 and 1999 as an alternative to
the MRTD and MRC and the procedure has essentially remained unchanged since except for some extensions and
adaptations for special cases such as: target motion, systems with noise drift or complex non-linear processing17,21,30.
Basically, a human observer judges the orientation (apex Up, Down, Left or Right) of equilateral triangle test patterns of
various sizes and contrasts using the sensor under test. See Figure 1. The threshold is defined at the 75% correct contrast
at a given angular size or the 75% correct angular size at a given contrast. A thorough statistical test has been
implemented to convert the observer response collection into the TOD curve: a threshold curve of contrast versus
reciprocal triangle angular size S-1 (see e.g. Figure 6, right picture). This threshold curve characterizes the quality of the
imaging system.
Over the years, a variety of TOD test equipment has been developed. These include: thermal test equipment13,30,34,42,43,
visual test charts and computer controlled test equipment8,21, an I2 test facility20, and an X-ray test31. Currently, the
demand for commercially available test equipment is growing and these are entering the market now.
Figure 1 The TOD test pattern: an equilateral triangle with apex Up, Down, Right or Left
2.1 Visual and NIR test charts
At TNO, high quality test charts for inside and outside use of camera testing in the visual and NIR spectral range were
developed in house. Both Visual Acuity (VA; effective resolution) and Contrast Sensitivity (CS) test charts exist with
positive and negative contrast test patterns (C = 1 for VA and C = 1 – 0 for CS). Background is grey with a reflectivity of
0.25 (near the average background on earth44) in the visible and near-infrared spectral region (see Figure 2). The charts
meet high requirements regarding 1) uniformity of test patterns and background, 2) the realization of high and low
contrast levels and 3) the edges of the targets. For image intensifier testing, additional requirements count for the spectral
properties in the visual and Near Infrared (NIR). They are calibrated for different standard light sources (D65, illuminant
A) and devices with different spectral sensitivity (visual, Gen II and Gen III image intensifiers). In addition, white charts
with black test patterns (C = -1) exist in accordance with the ECE-46 standard45 for testing visual devices in vehicles and
other applications.
Although the charts are primarily produced for own use, they can now be ordered by interested companies in limited
amounts (standard or custom made, see Figure 2, right picture).
Figure 2 TOD sensor test charts for lab and field testing; Left and middle picture: test chart for the measurement of
Visual Acuity (VA) and Contrast Sensitivity (CS) in the field. Right picture: production of test charts.
2.2 Thermal test bench
The first thermal test bench with TOD on the market is manufactured by HGH infrarouge46. A view on their user
interface is shown in Figure 3.
Figure 3. user interface of the TOD test in a commercially available IR test bench (image provided by HGH
Infrarouge)
2.3 Objective testing
Automated sensor testing using a validated objective synthetic human observer model is within reach. The HVS model
developed for TA modeling (see Section 3.3) can be applied to replace the human observer in an end-to-end system
performance test such as the TOD test in Section 2.2. This also allows more advanced applications such as the automatic
optimization of signal enhancement parameters36.
3. TOD MODELING
3.1 Standard perception database
For the purpose of model development, a standard (and extendable) TOD perception database has been developed for a
set of simulated sensor systems in which the primary sensor parameters (i.e. blur, sampling and noise) are systematically
varied. Such a set allows good control over the accuracy of the model predictions. Models that are based on perception
data for arbitrarily available combinations of parameter values may be tuned well to certain sensor systems (i.e. certain
positions in parameter space) or groups of sensor systems but there is always uncertainty about the predictive value for
other systems.
The current database47 consists of TOD thresholds for i) the unaided human eye and ii) a set of 24 simulated sensor
systems. The sensor systems have a fixed diffraction blur and a pixel pitch that is down-sampled by a decimation factor
of 2, 3, 4 and 8, respectively. With these decimation factors, the sensors in the set change from well-sampled to under-
sampled: (F·λ/d), i.e. a measure for the amount of under-sampling48, varies from 2 to 0.5. Further, static Gaussian noise
is added with σ = 0.001, 0.02 and 0.25. Up-sampling occurs in two different ways: pixel replication and bilinear
interpolation. In a recent study47, experiments were run at different observer viewing distances and display luminance
levels. For more details, we refer to Bijl et al. (2016, 2014)47,49. An extension of the database (e.g. towards dynamic
imaging and moving targets) is planned.
Example images of TOD test patterns imaged by the 24 sensors are shown in Figure 4 (pixel replication) and Figure 5
(bilinear interpolation). The TOD curves for part of this standard set (at the original display luminance L = 50 cd/m2 and
viewing distance) are given in Figure 6 for the unaided eye, and in Figure 7 for the simulated sensors. The top line
graphs in Figure 7 represent the low noise systems (0.001), middle line graphs represent medium noise (0.02) and the
bottom line shows the high noise (0.25) conditions. From left to right, sample size increases: decimation factor = 2, 3, 4
and 8 respectively. The results for different interpolation types (nearest neighbor: blue dots and bicubic: red dots) are
plotted together. A thorough analysis of the data is provided earlier (2014). In general, the high-contrast cut-off decreases
with increasing pixel pitch (as expected), contrast threshold increases with noise level (as expected), but the effect of
interpolation type is negligible although the images are very different (compare Figure 4 and Figure 5).
In Figure 8, all data from the 24 sensors are replotted together as threshold contrast /noise (= SNR, dimensionless) as a
function of pixel size d/test pattern size S (dimensionless) on a log-log scale. Thresholds for low noise sensors (n =
0.001) are plotted with open symbols, thresholds for medium noise in grey, and high noise (n = 0.25) in black symbols.
Different symbols indicate different decimation factors. Thresholds for nearest neighbor and bicubic interpolation are
plotted with identical symbols. The black line with slope 1 marks the lower visual threshold boundary induced by static
system noise. A slope of 1 is an interesting finding meaning that the total noise over the area of the test pattern is
independent of the size of the test pattern (expressed in pixels). Thresholds for the same noise level but different
decimation factor all fall on the same curve. The black symbols are purely noise limited, the grey symbols are limited by
both noise and human contrast sensitivity, and the open symbols are purely limited by human contrast sensitivity. The
vertical black line indicates the limit induced by sampling. This limit (d/S = 0.6) has been found in many TOD studies.
Figure 4 Example images of the TOD test patterns with the simulated sensors described in 3.1. Decimation factor
increases to the right (2,3,4, and 8). Top row: noise level 0.001; medium row: noise level 0.02; bottom row: noise level
0.25. Interpolation type: nearest neighbor (pixel replication).
Figure 5 Example images of the TOD test patterns with the simulated sensors described in 3.1. Decimation factor
increases to the right (2,3,4, and 8). Top row: noise level 0.001; medium row: noise level 0.02; bottom row: noise level
0.25. Interpolation type: bilinear.
0.001
0.01
0.1
1
0 0.5 1 1.5 2 2.5
Co
ntr
ast
1/S (mrad -1)
TOD naked eye, individual observers
FK
JA
MH
PB
WV
average
0.001
0.01
0.1
1
0 0.5 1 1.5 2 2.5
Co
ntr
ast
1/S (mrad -1)
TOD naked eye, L = 50 cd/m2
average
fit 50 cd/m2
Figure 6 TOD curves for the unaided human eye at a background luminance L = 50 cd/m2. Plotted is threshold contrast
against the reciprocal test pattern size S-1 or Visual Acuity (VA) in mrad-1. Left graph: the curves for the five observers
are plotted separately and the logarithmic mean over the observers (orange circle). Right: average over the observers and
a linear fit through the data.
0.01
0.1
1
0 0.2 0.4 0.6 0.8 1
Co
ntr
ast
1/size (mrad-1)
decimation: 2; noise: 0.001
nearest
bicubic
0.01
0.1
1
0 0.2 0.4 0.6 0.8
Co
ntr
ast
1/size (mrad-1)
decimation: 2; noise: 0.02
nearest
bicubic
0.001
0.01
0.1
1
0 0.2 0.4 0.6 0.8 1
Co
ntr
ast
1/size (mrad-1)
decimation: 3; noise: 0.001
nearest
bicubic
0.001
0.01
0.1
1
0 0.2 0.4 0.6 0.8 1
Co
ntr
ast
1/size (mrad-1)
decimation: 4; noise: 0.001
nearest
bicubic
0.01
0.1
1
0 0.2 0.4 0.6 0.8 1
Co
ntr
ast
1/size (mrad-1)
decimation: 8; noise: 0.001
nearest
bicubic
0.01
0.1
1
0 0.2 0.4 0.6 0.8 1
Co
ntr
ast
1/size (mrad-1)
decimation: 3; noise: 0.02
nearest
bicubic
0.01
0.1
1
0 0.2 0.4 0.6 0.8 1
Co
ntr
ast
1/size (mrad-1)
decimation: 4; noise: 0.02
nearest
bicubic
0.01
0.1
1
0 0.2 0.4 0.6 0.8 1
Co
ntr
ast
1/size (mrad-1)
decimation: 8; noise: 0.02
nearest
bicubic
0.01
0.1
1
0 0.2 0.4 0.6 0.8 1
Co
ntr
ast
1/size (mrad-1)
decimation: 2; noise: 0.25
nearest
bicubic
0.01
0.1
1
0 0.2 0.4 0.6 0.8 1
Co
ntr
ast
1/size (mrad-1)
decimation: 3; noise: 0.25
nearest
bicubic
0.01
0.1
1
0 0.2 0.4 0.6 0.8 1
Co
ntr
ast
1/size (mrad-1)
decimation: 4; noise: 0.25
nearest
bicubic
0.01
0.1
1
0 0.2 0.4 0.6 0.8 1
Co
ntr
ast
1/size (mrad-1)
decimation: 8; noise: 0.25
nearest
bicubic
Figure 7 TOD curves for the 24 simulated sensors. Blue dots interconnected by blue lines: nearest neighbor interpolation;
red dots interconnected by red lines: bicubic interpolation. The top line graphs: low noise (0.001); middle line graphs:
medium noise (0.02); bottom line graphs: high noise (0.25) conditions. From left to right: decimation factor = 2, 3, 4, 8.
Note that the contrast scale in the two top middle graphs is different from the others.
0.1
1
10
100
1000
0.01 0.1 1
Co
ntr
ast/
no
ise
= SN
R
pixel size/ triangle size
all sensors
dec = 2, n = 0.001dec = 3, n = 0.001dec = 4, n = 0.001dec = 8, n = 0.001dec = 2, n = 0.02dec = 3, n = 0.02dec = 4, n = 0.02dec = 8, n = 0.02dec = 2, n = 0.25dec = 3, n = 0.25dec = 4, n = 0.25dec = 8, n = 0.25Series13Series14Series15Series16Series17Series18Series19
rc = 1
Figure 8 TOD data for the 24 sensors from Figure 7 all together in a single plot, replotted as threshold contrast /noise
(dimensionless) as a function of pixel size d/test pattern size S (dimensionless) on a log-log scale. Open symbols: n =
0.001; grey symbols: n = 0.02 and black symbols: n = 0.25. Diamonds: decimation factor = 2; squares: decimation 3;
triangles: decimation 4, and circles: decimation 8. Thresholds for nearest neighbor and bicubic interpolation are plotted
with identical symbols. The black line with slope 1 marks the lower visual threshold boundary induced by static system
noise; the vertical black line indicates the limit induced by sampling. This limit (d/S = 0.6) has been found in many
TOD studies. See text for further details
3.2 Analytical TOD model
The analytical model presented here is an improvement of the earlier version derived by Bijl et al. (2016)47. A paper
describing the model in detail is in preparation50. Overall system performance is limited by four components:
(diffraction) blur, sampling, noise, and the TOD of the human observer viewing a TOD test pattern on the display.
All four are expressed as an equation of VA = S-1 (in mrad-1) as a function of test pattern contrast Cfield. The total
performance is the inverse Pythagorean sum of the components. The equations are:
VAhuman eye (Cfield) = Msys ∙[ a0 ∙ (1 + a1 ∙ Log [(Cfield∙α∙(naveragehuman /nhuman))])] (VA ≥ 0) (1)
VAblur (Cfield) = b0∙ (D/λ) ∙ (1+(a1∙Log[(Cfield∙α∙(naveragehuman /nhuman))])] (VA ≥ 0) (2)
VAsampling (Cfield) = (h/d) (3)
VAnoise (Cfield) = (h/d)• d1 • (Cfield/nsystem) (4)
VAtotal = [VAhuman eye-β + VAblur
-β + VAsampling-β + VAnoise-β] -1/β (5)
For a thermal imager, system noise is determined by :
nsystem (K) = √(IETD2 + (tframe/tint,eye) ∙ NETD2) (6)
All model parameters and values for the constants (valid for a display luminance L = 50 cd/m2) are listed in Table 1.
Table 1 TOD analytical model parameters (left table) and constants (right table)
parameter parameter name constant value
VA (mrad-1) Visual Acuity a0 (mrad-1) 1.48
Cfield Inherent target field contrast a1 , b1 0.52
Msys System angular magnification b0 1.9
α (in K-1 for thermal) System contrast multiplication d1 0.12
D (mm) Optics diameter h 0.65
λ(μm) Optics diffraction wavelength β 2
d (mrad) Angular detector pitch (IFOV) naveragehuman 0.04
nhuman 0.04 (default)
tint,eye (ms) 100
The prediction for the unaided eye is given by the orange line in Figure 6 (right-hand side picture), and Figure 9 gives
the predictions for the 24 sensor systems together with the data from Figure 7.
0.01
0.1
1
0 0.2 0.4 0.6 0.8 1
Co
ntr
ast
1/size (mrad-1)
decimation: 2; noise: 0.001
nearest
bicubic
model
0.01
0.1
1
0 0.2 0.4 0.6 0.8 1
Co
ntr
ast
1/size (mrad-1)
decimation: 2; noise: 0.02
nearest
bicubic
model
0.001
0.01
0.1
1
0 0.2 0.4 0.6 0.8 1
Co
ntr
ast
1/size (mrad-1)
decimation: 3; noise: 0.001
nearest
bicubic
model
0.001
0.01
0.1
1
0 0.2 0.4 0.6 0.8 1
Co
ntr
ast
1/size (mrad-1)
decimation: 4; noise: 0.001
nearest
bicubic
model
0.01
0.1
1
0 0.2 0.4 0.6 0.8 1
Co
ntr
ast
1/size (mrad-1)
decimation: 8; noise: 0.001
nearest
bicubic
model
0.01
0.1
1
0 0.2 0.4 0.6 0.8 1
Co
ntr
ast
1/size (mrad-1)
decimation: 3; noise: 0.02
nearest
bicubic
model
0.01
0.1
1
0 0.2 0.4 0.6 0.8 1
Co
ntr
ast
1/size (mrad-1)
decimation: 4; noise: 0.02
nearest
bicubic
model
0.01
0.1
1
0 0.2 0.4 0.6 0.8 1
Co
ntr
ast
1/size (mrad-1)
decimation: 8; noise: 0.02
nearest
bicubic
model
0.01
0.1
1
0 0.2 0.4 0.6 0.8 1
Co
ntr
ast
1/size (mrad-1)
decimation: 2; noise: 0.25
nearest
bicubic
model
0.01
0.1
1
0 0.2 0.4 0.6 0.8 1
Co
ntr
ast
1/size (mrad-1)
decimation: 3; noise: 0.25
nearest
bicubic
model
0.01
0.1
1
0 0.2 0.4 0.6 0.8 1
Co
ntr
ast
1/size (mrad-1)
decimation: 4; noise: 0.25
nearest
bicubic
model
0.01
0.1
1
0 0.2 0.4 0.6 0.8 1
Co
ntr
ast
1/size (mrad-1)
decimation: 8; noise: 0.25
nearest
bicubic
model
Figure 9 TOD curves for the 24 simulated sensors measured at Lbackground = 50 cd/m2 and at the original viewing
distance. Plotted together are: i) the data from Figure 7 (blue and red dots) and ii) predictions with the first order
analytical model of section 3.2 (black lines).
Using equation (5) for the system noise, predictions can be made for thermal imaging systems. In
Figure 10, model predictions are made for historical data published in earlier papers. Left graph: TOD for the simulated
‘star’ staring MWIR sensor system (from Hogervorst et al., 2001)37. Note that we discovered an error in the calculation
of ΔT: the factor k in equation 4 of the original paper should be 1/1.75 instead of 1.75. If we apply this factor, the
agreement is excellent. Right graph: TOD for the Indigo Omega uncooled microbolometer measured by Kostrzewa et al
(2003). Again, the data are predicted well.
0.01
0.1
1
0 5 10 15 20
∆T
(K)
1/size (mrad-1)
STAR sensor
data MAH
data PB
VA total (mrad-1)
VA human eye
VA diffraction
VA s + n
FNyquist
0,1
1
0 0,1 0,2 0,3 0,4
∆T
(K)
1/size (mrad-1)
Indigo Omega (2003)
VA total (mrad-1)
VA human eye
VA diffraction
VA s + n
FNyquist
data
Figure 10 Comparison of the analytical model predictions with historical data. Left graph. TOD for the simulated
‘star’ staring MWIR sensor system (from Hogervorst et al., 2001)37. Open and filled symbols: observer data. Blue
line: overall prediction, based on the component VAhuman eye (red), VAdiffraction (green) and VAsampling+noise (purple).
The agreement is excellent. Right graph. TOD for the Indigo Omega uncooled microbolometer measured by
Kostrzewa et al (2003)30. Dash-dot green line: replot from their Figure 10. Blue curve: model prediction. Again,
the data are predicted well.
3.3 Image-based Human Visual System model for TOD applications
We developed a practical static Human Visual System (HVS) model that can be applied to simulated imagery (for TA
modeling) or to the images from a sensor under test (for objective sensor testing). A third application is the optimum
system design trade-off between Size, Weight, Power and Cost (SWPaC) on the one hand, and performance on the other
The model consists of several parts: i) an algorithm to find the center of an object in an image, ii) a model to find the
most probable orientation by correlating the sensor output image with input images containing a test target with one of
the four possible test pattern orientations, and iii) the probability versus size and contrast function for a human observer
viewing the test pattern image on a display. In this way, the sensor-, display- and human-related factors are included such
as optics blur, sampling, noise, e-zoom, signal processing, display size, luminance level, viewing distance, observer
acuity and contrast sensitivity limitations.
Figure 11 shows the predicted model thresholds (blue dots) together with the human perception data (red dots) for the 12
sensors from the standard database (section 3.1) with nearest neighbor interpolation. The filled blue dots are valid
threshold estimates (maximum likelihood contrast thresholds fall within the range of presented stimuli and are approved
on the basis of a chi-square statistic), the open blue circles show the thresholds that are not statistically sound or fall
outside the range of available thresholds. The model shows a good overall agreement and there are no systematic
deviations on the basis of the amount of under-sampling (from left to right) or the amount of system noise (from top to
bottom).
Total Error=2.1718
Error3=1.3499
Error2=0.47488
Error1=0.347
Threshold=invweighted-0.0075
2: Human
1: correlation model
o: threshold OoB
interpolation: nearest
0 0.2 0.4 0.6 0.8 110
-4
10-2
100
Decimation: 2, Noise: 0.001, P-effect: -16 -4 -6 0
1/Size (1/mrad)
Contr
ast
0 0.2 0.4 0.6 0.8 110
-4
10-2
100
Decimation: 3, Noise: 0.001, P-effect: -23 -8 -2 0
1/Size (1/mrad)
Contr
ast
0 0.2 0.4 0.6 0.8 110
-4
10-2
100
Decimation: 4, Noise: 0.001, P-effect: -16 -13 -4 0
1/Size (1/mrad)
Contr
ast
0 0.2 0.4 0.6 0.8 110
-4
10-2
100
Decimation: 8, Noise: 0.001, P-effect: -15 -25 -13 0
1/Size (1/mrad)
Contr
ast
0 0.2 0.4 0.6 0.8 110
-4
10-2
100
Decimation: 2, Noise: 0.02, P-effect: -6 -2 -1 0
1/Size (1/mrad)
Contr
ast
0 0.2 0.4 0.6 0.8 110
-4
10-2
100
Decimation: 3, Noise: 0.02, P-effect: -3 0 0 0
1/Size (1/mrad)
Contr
ast
0 0.2 0.4 0.6 0.8 110
-4
10-2
100
Decimation: 4, Noise: 0.02, P-effect: -2 0 0 0
1/Size (1/mrad)
Contr
ast
0 0.2 0.4 0.6 0.8 110
-4
10-2
100
Decimation: 8, Noise: 0.02, P-effect: 0 0 0 0
1/Size (1/mrad)
Contr
ast
0 0.2 0.4 0.6 0.8 110
-4
10-2
100
Decimation: 2, Noise: 0.25, P-effect: 0 0 0 0
1/Size (1/mrad)
Contr
ast
0 0.2 0.4 0.6 0.8 110
-4
10-2
100
Decimation: 3, Noise: 0.25, P-effect: 0 0 0 0
1/Size (1/mrad)
Contr
ast
0 0.2 0.4 0.6 0.8 110
-4
10-2
100
Decimation: 4, Noise: 0.25, P-effect: 0 0 0 0
1/Size (1/mrad)
Contr
ast
0 0.2 0.4 0.6 0.8 110
-4
10-2
100
Decimation: 8, Noise: 0.25, P-effect: 0 0 0 0
1/Size (1/mrad)
Contr
ast
Figure 11 TOD data for the 12 simulated sensors with nearest neighbor interpolation (see section 3.1). Red dots:
observer data. Blue dots: Predictions using the Human Visual System (HVS) model described in section 3.3. The
agreement is good for all types of sensors in the set. See text for details.
4. TA RANGE MODELING
4.1 Conventional TA range modeling
The equations for conventional modeling are provided in Bijl & de Vries (2010)8. Basically, this is the ACQUIRE51
model in which the original Johnson criterion N50 has been replaced by a M75, i.e. the ratio of the 75% correct
characteristic size of the object set and that of the TOD test pattern which is associated with task difficulty. In 2011, a
comparison study (Bijl 2011) on the perception data underlying the TTP metric and the TOD has led to a conversion
factor between V50 and M75, enabling the exchange of US and NL target perception studies for the two models.
4.2 New approach
In a recent study35, threshold ranges for ship detection and for classification of characteristic ship parts (accommodation,
hull, mast, funnel, hat) were determined with the unaided human eye under a wide variety of conditions with different
sun directions, cloudiness, amount of atmospheric reduction and sea reflection. A few conditions are presented in Figure
12. In total, 1584 ranges were measured. Next, the ship contour, ship parts and corresponding TA ranges were analyzed
applying several feature size and luminance contrast measures. The size and contrast measure yielding the lowest
residual variance in the contrast versus threshold angular size relationship were identified as the best characteristic
Target Acquisition size and contrast measure. Finally, the TOD curve for the unaided eye was plotted together with the
contrast versus characteristic size data.
The most important results were:
The best agreement between data across target features was found when we defined the long axis of the feature as
the characteristic size of the feature and local contrast as the characteristic contrast.
The TOD curve practically coincides with the ship detection and feature classification data, i.e. M75 = 1 for ship
detection and feature classification
The results suggest that ship detection and the classification of ship features can be predicted by replacing those features
by a TOD pattern of the same characteristic size and local contrast. If the hypothesis is true, the results may directly be
applied to assess the effect on DRI ranges of features that distinct a target from potential alternatives. It also brings the
second stage of TA models to the next level where spatial target information is included in the range modeling.
Implementation into the existing TA models is simple: V50 or M75 for any target set may be assessed by indicating the
features that distinct one target from the others
Figure 12 Some examples of the images used in the ship detection and classification experiment (from Bijl et al,
2016).
5. APPLICATIONS
5.1 European Target Acquisition model ECOMOS
Both the analytical model (section 3.2) and an image-based TOD model are currently being implemented in the
European Target Acquisition Model ECOMOS40. ECOMOS is a multinational effort within the framework of an EDA
Project Arrangement. Its aim is to provide a generally accepted and harmonized European computer model for
computing nominal Target Acquisition (TA) ranges of optronic imagers operating in the Visible or thermal Infrared (IR).
The project involves close co-operation of Defence and security industry and public research institutes from France,
Germany, Italy, The Netherlands and Sweden. ECOMOS uses and combines well-accepted existing European tools to
build up a strong competitive position. This includes two TA models: the analytical TRM4 model and the image-based
TOD model. In addition, it uses the atmosphere model MATISSE.
The TOD branch consists of the following elements:
The analytical model in order to get a first order estimate of the expected performance and do an efficient Monte
Carlo simulation in the image-based calculation.
A simulation of the test procedure including the generation of TOD test patterns.
A simulation of the sensor system.
A plug-in option to include advanced signal processing if desired.
A simulation of the display.
The HVS model (see section 3.3).
The TOD TA range prediction model (as described in section 4.1).
An extensive description of ECOMOS is provided is a separate paper by Repasi et al. (2017)40.
5.2 Implementation in EOSTAR TDA
The Electro-Optical Signal Transmission and Ranging (EOSTAR) model is an image-based Tactical Decision Aid
(TDA) for thermal imaging systems (MWIR/LWIR) developed for a sea and littoral environment with an extensive
atmosphere model. In a current effort41, the TOD method is implemented in EOSTAR. Ship representative TOD test
patterns are placed at the position of the real target; the combined effects of the environment (atmosphere, background,
etc.), sensor and signal processing on the image are calculated using EOSTAR (see Figure 13 for an example), and the
results are judged using the HVS model. The thresholds are then converted into Detection-Classification-Identification
(DCI) ranges of the real target. This allows the prediction of Target Acquisition (TA) performance over the exact path
from scene to observer.
Figure 13 TOD simulation in EOSTAR, allowing the prediction of Target Acquisition (TA) performance over the exact
path from scene to observer. Ship representative TOD test patterns of different contrasts and orientations are positioned
at different ranges in the scene. The combined effects of the environment and the sensor are calculated using EOSTAR
and performance is assessed using the HVS and TOD TA range model.
6. CONVERSIONS BETWEEN TOD, TRM4 AND TTP
In earlier studies45,52,53 we showed that there are systematic system-dependent differences in the TA predictions made by
different models including the TOD model, TRM4 and TTP metric. For example, the MTDP cut-off crosses the TOD
cut-off when the sensor system changes from under-sampled to well-sampled45. Understanding the differences is
important for instance for the STANAG 4347 update that contains all three models and two tests. The TOD perception
dataset and the analytical model presented in Chapter 3 allow a systematic investigation of these differences.
In Figure 14, we plot MTDP curves calculated with TRM4 together with the perception data and the analytical model
predictions from Figure 9. In these plots, bar pattern spatial frequency in cy/mrad is plotted as 1/size, i.e. 1 cycle
corresponds with the TOD triangle square-root area (in mrad).
At low spatial frequencies, the MTDP contrast thresholds (i.e. thermal contrasts in this case) are much lower than the
TOD, especially for the low noise conditions (top graphs). In addition, the ratio between MTDP and TOD cut-off
systematically increases with decimation factor, i.e. with the amount of under-sampling in the system.
Studying the differences between the TRM4/MTDP and TOD, we made the following observations:
TRM4 does not include a human contrast sensitivity limiting model and assumes maximum contrast of the test
patterns at low spatial frequencies (i.e. always system noise limited). When measuring the MTDP, it is allowed to
adjust display level with the effect that contrast to the observer is increased and noise becomes the limiting factor.
For the same spatial frequency, the size of the four bars in the bar test pattern is 7 times larger than the area of the
TOD triangle. In a SNR limited region, this will lead to a (see Figure 8) threshold reduction by a factor of √7.
Note that these findings only hold for conditions where observer viewing distance plays an important role.
We take these factors into account by i) modeling the TOD for slightly adapted sensor systems with α = ΔTfield-1, leading
to a test pattern that always has a contrast Cdisplay = 1 at the display, and ii) correction of the TRM4 contrast thresholds by
a factor of √7. Replotting the data in terms of SNR as a function of d/size (similar to Figure 8), all on same scales, results
in Figure 15. We see that, in this representation:
TOD and MTDP thresholds at low spatial frequencies are very close to each other. Minor residual differences may
be ascribed to differences in modeling and measurement procedures, the particular choice of SNR in TRM4, etc.
The differences at low spatial frequencies are independent of noise level (from top to bottom graphs) and of
decimation factor (from left to right).
When the thresholds are plotted as SNR, the graphs for all noise levels (from top to bottom) fall exactly on top of
each other. We conclude that SNR, and not contrast, is the primary factor
There is a residual effect of (F·λ/d): for well-sampled systems the predicted MTDP cuts off before the TOD, there is
a cross-over at the transition area and for under-sampled systems the MTDP cuts off above the TOD. This is exactly
the effect described earlier by Bijl et al. (2003, 2016)45,52.
Summarizing
If human Contrast Sensitivity plays no role, then the ratio between limiting VA and limiting MTDP spatial
frequency is a function of both (F·λ/d) and SNR. This is probably a sigmoidal function: the ratio is constant at well-
sampled systems and for highly under-sampled systems and varies near the transition area45. Of course, this is a first
order approximation because only the major system parameters are taken into account.
Conversion from MTDP to TOD can be done as follows: MTDP contrast thresholds are multiplied by a factor of √7.
Then, the conversion factor f [(F·λ/d), SNR] is applied to MTDP spatial frequency. If the contrast of the target on
the display is lower than 1, then a correction for human-limited CS on VA needs to be performed using equation (1).
Finally, for range calculation, M75 should be applied instead of N50.
A more detailed analysis and validation is possible when ECOMOS becomes available. Similar comparison studies can
be performed between TTP and TOD and/or TTP and TRM4.
0,0001
0,001
0,01
0,1
1
10
0 0,2 0,4 0,6 0,8 1
Co
ntr
ast
1/size (mrad-1)
decimation: 2; noise: 0.001
nearest
bicubic
TOD
TRM4
0,001
0,01
0,1
1
10
0 0,2 0,4 0,6 0,8 1
Co
ntr
ast
1/size (mrad-1)
decimation: 2; noise: 0.02
nearest
bicubic
TOD
TRM4
0,0001
0,001
0,01
0,1
1
10
0 0,2 0,4 0,6 0,8 1
Co
ntr
ast
1/size (mrad-1)
decimation: 3; noise: 0.001
nearest
bicubic
TOD
TRM4
0,0001
0,001
0,01
0,1
1
10
0 0,2 0,4 0,6 0,8 1
Co
ntr
ast
1/size (mrad-1)
decimation: 4; noise: 0.001
nearest
bicubic
TOD
TRM4
0,0001
0,001
0,01
0,1
1
10
0 0,2 0,4 0,6 0,8 1
Co
ntr
ast
1/size (mrad-1)
decimation: 8; noise: 0.001
nearest
bicubic
TOD
TRM4
0,001
0,01
0,1
1
10
0 0,2 0,4 0,6 0,8 1
Co
ntr
ast
1/size (mrad-1)
decimation: 3; noise: 0.02
nearest
bicubic
TOD
TRM4
0,001
0,01
0,1
1
10
0 0,2 0,4 0,6 0,8 1
Co
ntr
ast
1/size (mrad-1)
decimation: 4; noise: 0.02
nearest
bicubic
TOD
TRM4
0,001
0,01
0,1
1
10
0 0,2 0,4 0,6 0,8 1
Co
ntr
ast
1/size (mrad-1)
decimation: 8; noise: 0.02
nearest
bicubic
TOD
TRM4
0,001
0,01
0,1
1
10
0 0,2 0,4 0,6 0,8 1
Co
ntr
ast
1/size (mrad-1)
decimation: 2; noise: 0.25
nearest
bicubic
TOD
TRM4
0,001
0,01
0,1
1
10
0 0,2 0,4 0,6 0,8 1
Co
ntr
ast
1/size (mrad-1)
decimation: 3; noise: 0.25
nearest
bicubic
TOD
TRM4
0,001
0,01
0,1
1
10
0 0,2 0,4 0,6 0,8 1C
on
tras
t1/size (mrad-1)
decimation: 4; noise: 0.25
nearest
bicubic
TOD
TRM4
0,001
0,01
0,1
1
10
0 0,2 0,4 0,6 0,8 1
Co
ntr
ast
1/size (mrad-1)
decimation: 8; noise: 0.25
nearest
bicubic
TOD
TRM4
Figure 14 Measured (blue and red symbols) and modeled TOD curves (black lines) for the simulated sensors
replotted from Figure 9, together with the MTDP as predicted by TRM4 (green lines). At low spatial frequencies,
the MTDP contrast thresholds (i.e. thermal contrasts in this case) are much (5-100x) lower than the TOD,
especially for the low noise conditions (top graphs). In addition, the ratio between MTDP and TOD cut-off
systematically increases with decimation factor, i.e. with the amount of under-sampling in the system.
0,1
1
10
100
1000
10000
0 0,2 0,4 0,6 0,8 1
SNR
d/size
decimation: 2; noise: 0.001
TOD
TRM4
0,1
1
10
100
1000
10000
0 0,2 0,4 0,6 0,8 1
SNR
d/size
decimation: 2; noise: 0.02
TOD
TRM4
0,1
1
10
100
1000
10000
0 0,2 0,4 0,6 0,8 1
SNR
d/size
decimation: 3; noise: 0.001
TOD
TRM4
0,1
1
10
100
1000
10000
0 0,2 0,4 0,6 0,8 1
SNR
d/size
decimation: 4; noise: 0.001
TOD
TRM4
0,1
1
10
100
1000
10000
0 0,2 0,4 0,6 0,8 1
SNR
d/size
decimation: 8; noise: 0.001
TOD
TRM4
0,1
1
10
100
1000
10000
0 0,2 0,4 0,6 0,8 1
SNR
d/size
decimation: 3; noise: 0.02
TOD
TRM4
0,1
1
10
100
1000
10000
0 0,2 0,4 0,6 0,8 1
SNR
d/size
decimation: 4; noise: 0.02
TOD
TRM4
0,1
1
10
100
1000
10000
0 0,2 0,4 0,6 0,8 1
SNR
d/size
decimation: 8; noise: 0.02
TOD
TRM4
0,1
1
10
100
1000
10000
0 0,2 0,4 0,6 0,8 1
SNR
d/size
decimation: 2; noise: 0.25
TOD
TRM4
0,1
1
10
100
1000
10000
0 0,2 0,4 0,6 0,8 1
SNR
d/size
decimation: 3; noise: 0.25
TOD
TRM4
0,1
1
10
100
1000
10000
0 0,2 0,4 0,6 0,8 1
SNR
d/size
decimation: 4; noise: 0.25
TOD
TRM4
0,1
1
10
100
1000
10000
0 0,2 0,4 0,6 0,8 1
SNR
d/size
decimation: 8; noise: 0.25
TOD
TRM4
Figure 15 TOD curves for the simulated sensors (black lines), but now predicted for noise limited
conditions. This was achieved by modeling a test pattern contrast of 1 on the observer display. TRM4
contrasts are shifted upwards by a factor of √7 to account for the difference in test pattern area. The data are
plotted as SNR as a function d/size. See text for further details.
7. CLOSING REMARKS
In the nineties of the last century, when it became clear that the standard MRTD/MRC and Johnson-based approach did
not apply to the new generation of staring imaging systems, a variety of alternatives appeared54-56. Three approaches
survived to date and are included in the draft updated STANAG for IR systems performance characterization.
Of these three approaches, the MTDP was a minor adaptation of the MRTD, allowing the characterization of staring and
conventional imaging systems. Since then, effort has been spent in order to formulate the sensor performance model and
to develop the extensive and robust TRM4 model software.
The TTP evolved from a coarse model (and software: NVTherm, NVThermIP) based on the perception of sinusoidal test
patterns and has continuously been improved and tuned to perception data up to a relatively stable model and robust and
modern software package called NV-IPM. A corresponding end-to-end sensor test has not been developed.
The TOD was built up from scratch getting the basics right and thus maximizing the chance that it is representing sensor
performance: a test that is close to the real acquisition task, a test pattern representing features of the targets that need to
be distinguished, a bias-free forced-choice observer task, and a solid statistical procedure. Next, the method has been
validated in many studies. Over the years, the TOD proved to be robust against the increasing system complexity.
Further, there is no fundamental limit to extend the method towards more complex and/or natural situations, e.g. using
motion, a complex background or a contrast higher than 1. With the ever increasing complexity of imaging systems,
image-based modeling has the future. It is expected that imaging simulation in general opens the field for a much wider
application of sensor and signal processing assessment techniques, and the TOD method forms a firm starting point.
REFERENCES
[1] Bijl, P.& Valeton, J.M. (1998). TOD, the alternative to MRTD and MRC. Optical Engineering 37, 7, 1976 - 1983.
[2] Wittenstein, W. (1999). Minimum temperature difference perceived – a new approach to assess under-sampled
thermal imagers. Optical Engineering 38, 5, 773 – 781.
[3] Vollmerhausen, R.H., Jacobs, E., Hixson, J. & Friedman, M. (2006). The Targeting Task Performance (TTP)
Metric: A New Model for Predicting Target Acquisition Performance. Technical Report AMSEL-NV-TR-230,
NVESD, Fort Belvoir, VA 22060 (January 2006).
[4] Vollmerhausen, Richard H., Eddie Jacobs, and Ronald Driggers (2004), New metric for predicting target acquisition
performance. Optical Engineering, Vol.43, No.11, pp 2806-2818.
[5] STANAG 4347:1995. Definition Of Nominal Static Range Performance For Thermal Imaging Systems. NATO
Standard Agreement (cancelled)
[6] Bijl, P.& Valeton, J.M. (1998). Validation of the new TOD method and ACQUIRE model predictions using
observer performance data for ship targets. Optical Engineering 37, 7, 1984 - 1994.
[7] Bijl, P.& Valeton, J.M. (1999). Guidelines for accurate TOD measurement. SPIE Proceedings, Vol. 3701 14 - 25.
[8] Bijl, P. & Vries, S.C. de, (2010). Visual Acuity, Contrast Sensitivity and Range Performance with compressed
motion video, Optical Engineering, Vol. 49, 103203, doi:10.1117/1.3503950.
[9] Bijl, P., Reynolds, J.P., Vos, W., Hogervorst, M.A. & Fanning, J.D. (2011). TOD to TTP calibration. In: Infrared
Imaging Systems: Design, Analysis, Modeling, and Testing XXII, 8014, 80140L; doi:10.1117/12.887219.
[10] Bijl, P., Hogervorst, M.A. & Vos, W. (2011). Direct comparison of TOD, TTP, and tactical vehicle identification
performance data underlying the TTP metric. (Report TNO-DV 2011 C486). Soesterberg, The Netherlands: TNO
Behavioural and Societal Sciences.
[11] Bijl, P. (2001). Validation of the TOD and MTDP Sensor Performance Measures for staring and scanning thermal
imagers (Report TNO-TM-01-A020). Soesterberg, The Netherlands: TNO Human Factors.
[12] Krapels, K., Driggers, R.G. & Teaney, B. (2005). Target-acquisition performance in under-sampled infrared
imagers: static imagery to motion video. Applied Optics, 44 (33), 7055-7061.
[13] Driggers, R.G., Krapels, K., Murrill, S., Young, S.S., Thielke, M. & Schuler, J. (2004). Superresolution performance
for undersampled imagers. Optical Engineering 44, 1.
[14] Jonathan Fanning, Justin Miller, Jennifer Park, Gene Tener, Joseph Reynolds, Patrick O’Shea, Carl Halford, Ron
Driggers (2007). IR system field performance with super-resolution. Proc. SPIE, Vol. 6543.
[15] Bijl, P., Beintema J. A., Dijk, J. & Leden, N. van der. (2010). Effectiveness assessment of signal processing in the
presence of smear. In: Infrared Imaging Systems: Design, Analysis, Modeling, and Testing XXI, 7662, 76620E-
76620E-10.
[16] Bijl, P. (2010). Visual Image Quality Assessment with sensor motion: Effect of Recording and Presentation
Velocity. Applied Optics, Vol. 49, Issue 3, pp. 343-349.
[17] Bijl, P., Schutte, K. & Hogervorst, M.A. (2006). Applicability of TOD, MRT, DMRT and MTDP for dynamic
image enhancement techniques. SPIE Proceedings 6207, 154-165.
[18] Bijl, P., Valeton, J.M. & de Jong, A.N. (2000). TOD predicts target acquisition performance for staring and
scanning thermal imagers, SPIE Proceeding Vol. 4030, 96-103.
[19] Bijl, P., Hogervorst, M.A. & Toet, A. (2004). Identification of military targets and simple laboratory test patterns in
band-limited noise. SPIE Proceedings Vol. 5407, 15, 104-115.
[20] Laurent, N., Bijl, P. & Deltel, G. (2015). Performance characterization of night vision equipment based on the
Triangle Orientation Discrimination (TOD) methodology. Optical Engineering 54(2), 023014.
[21] Gosselink, G.A.B., Anbeek, H., Bijl, P. & Hogervorst, M.A. (2013). TOD Characterization of the Gatekeeper
Electro Optical Security System. In: Proc. SPIE 8706, Infrared Imaging Systems: Design, Analysis, Modeling, and
Testing XXIV, 87060I; doi:10.1117/12.2016589.
[22] Dijk, J., Eekeren, A.W.M. van, Toet, A., Hollander, R.J.M. den, Schutte, K., Heijningen, A.W.P. van, Bijl, P.
(2013). Evaluation of intensified image enhancement through conspicuity and TOD measures. Opt. Eng. 52 (4),
041105 (December 04, 2012); doi: 10.1117/1.OE.52.4.041105.
[23] Desaulniers, P. & Thibault, S. (2011). Performance evaluation of panoramic electro-optic imagers using the TOD
method. In: Infrared Imaging Systems: Design, Analysis, Modeling, and Testing XXII, 8014-801409;
doi:10.1117/12.883718
[24] Gaska, J.P., Winterbottom, M., Sweet, W., Rader, J. (2010). Pixel size requirements for eye-limited flight
simulation. Image 2010 conference, July 2010, Scottsdale, Arizona.
[25] Dijk, J., Bijl, P., Eekeren, A.W.M. van. (2010). Performance evaluation of image enhancement techniques on an
EMCCD camera. In: Infrared Imaging Systems: Design, Analysis, Modeling, and Testing XXI, 7662, 766207-
766207-12.
[26] Eekeren, W.M. van, Schutte, K., Oudegeest, O.R. & Vliet, L.J. van (2007). Performance Evaluation of Super-
Resolution Reconstruction Methods on Real-World Data. EURASIP Journal on Advances in Signal Processing, vol.
2007, Article ID 43953.
[27] S. Hu, S. S. Young, T. Hong, J.P. Reynolds, K. Krapels, B. Miller, J. Thomas, and O. Nguyen (2010). Super-
resolution for flash ladar imagery. Applied Optics, Vol. 49, Issue 5, pp. 772-780
[28] Keith Krapels, Ronald Driggers, Eddie Jacobs, Stephen Burks, and Susan Young (2007). Characteristics of Infrared
Imaging Systems which benefit from Super-Resolution Reconstruction. Applied Optics, Vol. 46, Issue 21, pp. 4594-
4603.
[29] Kooi, F.L., Bijl, P., & Padmos, P (2006). Stereo Acuity and Visual Acuity in Head Mounted Displays. Proc. Of the
Human Factors and Ergonomics Society, 2693-2696.
[30] Kostrzewa, J., Long, J., Graff, J. & Vincent, J.D. (2003). TOD versus MRT when evaluating thermal imagers that
exhibit dynamic performance. SPIE Proceedings Vol. 5076, 220 – 232.
[31] Bijl, P., Hogervorst, M.A., Valeton, J.M. & Ruiter, C.J. de (2003). BAXSTER: An Image Quality Tester for X-ray
Baggage Screening Systems. SPIE Proceedings Vol. 5071, 341-352.
[32] Gunnar Ritt ; Michael Koerber ; Daniel Forster ; Bernd Eberle (2015). Protection performance evaluation regarding
imaging sensors hardened against laser dazzling. Opt. Eng. 54(5), 053106 (May 08, 2015).
doi:10.1117/1.OE.54.5.053106.
[33] Alan R. Pinkus, A.R., Dommetta, D.W. & Task, H.L (2012). A Comparison of Landolt C and triangle resolution
targets using the synthetic observer approach to sensor resolution assessment. In: Proc. SPIE 8355, Infrared
Imaging Systems: Design, Analysis, Modeling, and Testing XXIII, 835515 (May 18, 2012); doi:10.1117/12.919225.
[34] Weiß, A. R., Adomeit, U., Chevalier, Ph., Landeau, S., Bijl, P., Champagnat, F., Dijk, J., Göhler, B., Landini, S.,
Reynolds, J.P., Smith, L. (2012). A standard data set for performance analysis of advanced IR image processing
techniques. In: Infrared Imaging Systems: Design, Analysis, Modeling, and Testing XXIII, Volume 8355, pp.
835512-835512-10.
[35] Bijl, P., Toet, A. & Kooi, F.L. (2016). Feature long axis size and local luminance contrast determine ship Target
Acquisition performance: strong evidence for the TOD case. Proc. SPIE 9987, Electro-Optical and Infrared Systems:
Technology and Applications XIII, 99870M (October 21, 2016); doi:10.1117/12.2241805
[36] De Lange, D.J., Valeton, J.M. & Bijl, P. (2000). Automatic characterization of electro-optical sensors with image-
processing, using the Triangle Orientation Discrimination (TOD) method. SPIE Proceedings, Vol. 3701, 104-111.
[37] Hogervorst, M.A., Bijl, P. & Valeton, J.M. (2001). Capturing the sampling effects: a TOD sensor performance
model. SPIE Proceedings Vol. 4372, 62-73.
[38] Bijl, P., Hogervorst, M.A. & Vos, W. (2008). Modular Target Acquisition model & visualization tool. In: Infrared
Imaging Systems: Design, Analysis, Modeling, and Testing XIX, 6941, 69410E.
[39] Labarre, L., Bijl, P., Repasi, E., Wittenstein, W. & Bürsing, H. (2016). The European COmputer Model for
Optronic System performance prediction (ECOMOS). OPTRO-2016-046 (France).
[40] Repasi, E., Bijl, P., Labarre, L., Wittenstein, W. & Bürsing, H. (2017). The European Computer Model for
Optronic System Performance Prediction (ECOMOS): Overview and State of Work. SPIE paper 10178-23 (in
press).
[41] Iersel, M. van., Bijl, P., Hove, R.J.M. ten., &Oppeneer, M., (2017). Target Acquisition modeling over the exact
optical path: Extending the EOSTAR TDA with the TOD sensor performance model. Abstract SPIE Security +
Defence (Europe) 2017.
[42] Valeton, J.M., Bijl, P., Agterhuis, E. & Kriekaard, S. (2000). T-CAT, a new Thermal Camera Acuity Tester. SPIE
Proceedings Vol. 4030, 232 – 238.
[43] Wang, C., Guo, X., Ren, T. & Zhang, Z. (2014). Performance evaluation of infrared imaging system in field test ",
Proc. SPIE 9300, International Symposium on Optoelectronic Technology and Application 2014: Infrared
Technology and Applications, 93002J (November 20, 2014); doi:10.1117/12.2074332;
http://dx.doi.org/10.1117/12.2074332.
[44] Bijl, P. (2016). How to pass a sensor acceptance test: using the gap between acceptance criteria and operational
performance. Proc. SPIE 9987, Electro-Optical and Infrared Systems: Technology and Applications XIII, 99870K
(October 21, 2016); doi:10.1117/12.2241792
[45] UN regulation ECE-46 (2012). Uniform provisions concerning the approval of devices for indirect vision and of
motor vehicles with regard to the installation of these devices (UN publication, 3 October 2012)
[46] HGH Infrared Systems. www.hgh.fr.
[47] Bijl, P. & Hogervorst, M.A. (2016). A first order analytical TOD sensor performance model. Proc. SPIE 9987,
Electro-Optical and Infrared Systems: Technology and Applications XIII, 99870P; doi: 10.1117/12.2242262
[48] G.C. Holst (2007). Imaging system performance based upon Fλ/d. Opt. Eng. 46(10), 103204 (October 03, 2007).
doi:10.1117/1.2790066.
[49] Bijl, P., Hogervorst, M.A. & Kooi, F.L. (2014). A TOD dataset to validate human observer models for target
acquisition modeling and objective sensor performance testing. Proc. SPIE 9249, Electro-Optical and Infrared
Systems: Technology and Applications XI, 92490R (October 20, 2014); doi:10.1117/12.2067454.
[50] Bijl, P. & Hogervorst, M.A. (2017). Analytical TOD sensor performance model. In preparation.
[51] Scott, L.B. & and Tomkinson, D.M. (1991). Update on the c2nveo FLIR90 and ACQUIRE sensor performance
models, 1991.
[52] Bijl, P., Hogervorst, M.A. & Valeton, J.M. (2002). TOD, NVTherm and TRM3 model calculations: a comparison.
SPIE Proceedings Vol. 4719, 51-62.
[53] Bijl, P. & Hogervorst, M.A.(2007). NVThermIP vs TOD: matching the Target Acquisition range criteria. SPIE
Proceedings 6543, pp. 65430C.
[54] Webb, C.M. (1994). "Minimum resolvable temperature difference--how far can we stretch it?", Proc. SPIE 2224,
Infrared Imaging Systems: Design, Analysis, Modeling, and Testing V, 297 (July 8, 1994); doi:10.1117/12.180068;
[55] Webb, C.M. &Halford, C.E. (1999). Dynamic minimum resolvable temperature difference testing for staring array
imagers. Opt. Eng., 38 (5) (1999), pp. 845–851
[56] Jong, A.N. de, Franken, E.M. & Winkel, H. (2003). Alternative measurement techniques for infrared sensor
performance. Opt. Eng. 42(3) 712–724.