Professional Documents
Culture Documents
4.highway Traffic Monitoring and Data Quality - (PG 86 - 111)
4.highway Traffic Monitoring and Data Quality - (PG 86 - 111)
4.highway Traffic Monitoring and Data Quality - (PG 86 - 111)
With so many factors at play, it is likely that enumerators will make mis-
takes and that variation is likely to be unpredictable.
If enumerators are to be assessed, they should be assessed either against a
video or other permanent test, or against each other in live situations. In the lat-
ter case, it is reasonable to assume initially that each enumerator is of the same
performance variance, and the net variance of each is assumed to be 1/√2 of the
difference between the two.
Unless further quantified information is at hand, a manual enumeration
error between 1% and 5% is not an unreasonable assumption.
site ID. This sounds like an obvious error, but it happens about 1 in a 100
times on temporary surveys.
• Technician is not following procedures or manuals or is inadequately
trained.
• A relevant quality manual and/or appropriate attitude by the employing
organization is lacking.
Dalgleish, Michael, and Neil Hoose. Highway Traffic Monitoring and Data Quality, Artech House, 2008. ProQuest Ebook
Central, http://ebookcentral.proquest.com/lib/utranscomm-ebooks/detail.action?docID=456896.
Created from utranscomm-ebooks on 2018-01-30 00:53:03.
Errors in Traffic Data 61
• Loops installed are not exactly square and central to lanes of traffic flow.
• Loops installed are the wrong size or different sizes when paired.
• Feeders are too long. Ideally feeders should be 50m or less, but sometimes
practical equipment location considerations make this impossible.
These error sources result partly from technological limitations and partly
from blunders. If they are a feature of a site, the documentation or file note for
the site should record the reasons.
Copyright © 2008. Artech House. All rights reserved.
1. The front and rear of the vehicle may be largely constructed of plastic, and
the distance from the main metal in the vehicle is variable from the phys-
ical edges.
2. The height of the metal at the front and rear of the vehicle varies
considerably.
Dalgleish, Michael, and Neil Hoose. Highway Traffic Monitoring and Data Quality, Artech House, 2008. ProQuest Ebook
Central, http://ebookcentral.proquest.com/lib/utranscomm-ebooks/detail.action?docID=456896.
Created from utranscomm-ebooks on 2018-01-30 00:53:03.
62 Highway Traffic Monitoring and Data Quality
The key point is that the loop detector measures the length of the con-
tiguous conductive area of the vehicle, not its true physical length. The con-
ductive length is a proxy for physical length but satisfactory for most traffic
applications.
The typical variation between the actual physical length and the actual
conductive length for cars is about ±20 cm at a 95% confidence level. The actual
errors in the detection technology need to be added to these values. This then
gives the gross error.
In other words, there are two types of error: proxy error, or the degree to
which the measurand is indicative of the parameter being measured, and detec-
tor error, or the error that the detector makes when measuring the proxy. The
upshot of all this for length measurement is that an optical technique is going to
get results much closer to the truth because the proxy error is so low.
Going back to the loop length, ±4 cm repeatability of length measurement
is routine for current loop detectors. Note this is not the accuracy but the repeat-
ability, or “precision,” of the repeated measurement of a single vehicle.
Microwave detectors also use more approximate methods to estimate
length, which will show much greater variation than loop detection.
used.
• Tubes that move about on the road surface, especially with large vehicles;
• The finite sampling resolution of the machine when measuring speed;
• Undercounting when vehicle speeds fall;
• Overcounting when vehicles stop or turn on the tubes and at high
speed.
Axle detectors also cause undercounting due to masking when placed over
two lanes. This may be modeled statistically. If tube detectors are used to count
Dalgleish, Michael, and Neil Hoose. Highway Traffic Monitoring and Data Quality, Artech House, 2008. ProQuest Ebook
Central, http://ebookcentral.proquest.com/lib/utranscomm-ebooks/detail.action?docID=456896.
Created from utranscomm-ebooks on 2018-01-30 00:53:03.
Errors in Traffic Data 63
• The age of the plate, as readings of old plates are more unreliable;
• Special and custom number plates, which will fail syntax checks more
frequently.
• There is heavy or very slow traffic, for example, during peak congested
hours;
Copyright © 2008. Artech House. All rights reserved.
• A vehicle in a near lane obstructs the view of vehicles in the target lane;
• A goods vehicle ahead obscures a vehicle behind;
• A goods vehicles behind obscures a vehicle in front;
• Lane changing avoids camera sensing.
In addition, site location factors affect the reading rate. Systems may be
installed,
Dalgleish, Michael, and Neil Hoose. Highway Traffic Monitoring and Data Quality, Artech House, 2008. ProQuest Ebook
Central, http://ebookcentral.proquest.com/lib/utranscomm-ebooks/detail.action?docID=456896.
Created from utranscomm-ebooks on 2018-01-30 00:53:03.
64 Highway Traffic Monitoring and Data Quality
The following environmental conditions can also affect the read rate:
• Direct or reflected sun light shining directly onto the camera glass;
• Bad weather, wind, precipitation, and/or bad visibility.
example, a system may ignore any plate that does not meet a fixed aspect ratio
between its width and height, so the percentage of successful reads will depend
on the proportion of the population of vehicles not conforming to the aspect
ratio. Environmental influences vary over time (e.g., position of the sun during
the day), so their effect on accurate plate reading will also vary overtime.
Most ANPR systems also need to record the time at which license plates are
captured. In contrast with the above, all ANPR have reliable and accurate clocks,
and it is relatively easy to check remotely for any time drift and correct it before
it becomes significant. For example, where journey times are being measured by
matching ANPR data from two locations, the significance of an error of a few
seconds depends on the distance between ANPR sites.
The significance of any bias in ANPR systems depends on the application
for which the data is collected. Enforcement applications need those plates that
Dalgleish, Michael, and Neil Hoose. Highway Traffic Monitoring and Data Quality, Artech House, 2008. ProQuest Ebook
Central, http://ebookcentral.proquest.com/lib/utranscomm-ebooks/detail.action?docID=456896.
Created from utranscomm-ebooks on 2018-01-30 00:53:03.
Errors in Traffic Data 65
are read to be correct but can accept a lower number of successful reads provided
that the credibility of the systems is not undermined. Journey time monitoring
needs consistent readings of plates at different sites to allow matches to be maxi-
mized, although the individual reads do not necessarily have to be correct.
Therefore, quantifying bias in ANPR means verifying its performance
against all vehicles that traverse the section during the verification period. This
entails a manual process in most cases. Note that due to the high prevalence of
time-varying sources of error, the selection of verification periods becomes im-
portant in designing a verification regime.
3.19 Summary
This chapter has focused on the systematic error rate and its determination. In
particular, it has highlighted a few critical data items:
This chapter has given worked examples to illustrate the differences be-
tween different assumptions and the associated calculations. Finally, the main
sources of error for a variety of data collection techniques have been discussed.
Copyright © 2008. Artech House. All rights reserved.
Dalgleish, Michael, and Neil Hoose. Highway Traffic Monitoring and Data Quality, Artech House, 2008. ProQuest Ebook
Central, http://ebookcentral.proquest.com/lib/utranscomm-ebooks/detail.action?docID=456896.
Created from utranscomm-ebooks on 2018-01-30 00:53:03.
4
Accuracy Assessments
4.1 Introduction
Up to now, we have made all our assessments in terms of systematic and ran-
dom error components. This is the preferred way to assess equipment. However,
some specifications refer to accuracy. Unfortunately, accuracy is not a universally
agreed upon term and has a number of different meanings.
In its simplest form, accuracy refers to the difference between a reported
measurement and the accepted reference value. For a number of reports, accura-
cy refers to either a confidence interval or a lack of systematic error or bias. While
these definitions sound straightforward, the data needed to support a claim of a
particular level of accuracy (e.g., ±1%) is more complicated. Most of this chapter
Copyright © 2008. Artech House. All rights reserved.
67
Dalgleish, Michael, and Neil Hoose. Highway Traffic Monitoring and Data Quality, Artech House, 2008. ProQuest Ebook
Central, http://ebookcentral.proquest.com/lib/utranscomm-ebooks/detail.action?docID=456896.
Created from utranscomm-ebooks on 2018-01-30 00:53:03.
68 Highway Traffic Monitoring and Data Quality
Each of these tests confirms that systematic errors or bias lie within a cer-
tain range with a certain confidence level.
Minimizing sample size is a key requirement to keep costs down and min-
imize time on-site. This chapter describes mathematical methods to calculate
the minimum sample size when determining compliance with accuracy speci-
fications. It also has been written to focus on counting accuracy, but much of
the material is common with continuous measurements such as vehicle speed,
length, and so forth. However, before starting that discussion, we need to deal
with the more difficult issue of combining the fixed, but unknown, systematic
error with the random error for an interval count.
or systematic error in the data produced by the TMS. This addressed the issue of
the long-term difference between true and reported counting. But, as the exam-
ple showed, the error in the TMS is not regular but varies from count to count.
On occasion, it is required to determine these random errors in interval
counts. For example, if tolls are being paid based on hourly traffic counts, then
the variation to be seen in hourly figures will be of interest in addition to the
mean error, because counts in one period may attract different charges to another
period.
Due to the random nature of miscounts, the random errors in small in-
tervals (e.g., 20 seconds or 1 minute) may be quite high compared with the
mean error rate. The random errors become more significant as the interval de-
creases but conversely approach zero for very long periods. These errors turn up
Dalgleish, Michael, and Neil Hoose. Highway Traffic Monitoring and Data Quality, Artech House, 2008. ProQuest Ebook
Central, http://ebookcentral.proquest.com/lib/utranscomm-ebooks/detail.action?docID=456896.
Created from utranscomm-ebooks on 2018-01-30 00:53:03.
randomly as the name suggests (sometimes in threes, like buses!), and it is this
variance for given intervals that we wish to quantify.
best estimate is –0.61% and that there is a 95% chance that it lies somewhere
between +0.13% and –1.35%. It may well lie near the middle of this interval,
but we don’t know that. So, we have to be impartial about where the mean might
actually be.
Now, if the mean error varied with every report, we would be justified in
combining it in quadrature with the random error. But since it doesn’t, the addi-
tion of the two deviations needs to be a straight arithmetic addition. Accepting
this approach means, in the case of continuing 10 minute counts with about 200
vehicles in each sample, that:
• In 95% of cases, the true count of vehicles will lie within (–0.61% –
2.13%) to (–0.61% + 2.13%) (i.e., –2.74% to +1.48%) of what was
reported.
Dalgleish, Michael, and Neil Hoose. Highway Traffic Monitoring and Data Quality, Artech House, 2008. ProQuest Ebook
Central, http://ebookcentral.proquest.com/lib/utranscomm-ebooks/detail.action?docID=456896.
Created from utranscomm-ebooks on 2018-01-30 00:53:03.
70 Highway Traffic Monitoring and Data Quality
• If all counts are adjusted by being multiplied by 1.0061, they will then be
accurate to ±2.13% with a confidence level of 95%.
CII n
CII m =
m (4.1)
n
For example, the CII for 10 minute intervals in Section 4.3 was 2.13%. To
calculate the CII for 60 minute intervals, use m = 60 and n = 10:
CII 10 ±2.13%
CII 60 = = = ±0.87%
60 2.45
10
Copyright © 2008. Artech House. All rights reserved.
Thus, the confidence interval for all individual 60 minute count reports is
estimated as
It is probably obvious that as the time period expands toward infinity, then
the confidence interval approaches that set for the mean. The confidence interval
for individual counts for any period can be determined using this method.
Dalgleish, Michael, and Neil Hoose. Highway Traffic Monitoring and Data Quality, Artech House, 2008. ProQuest Ebook
Central, http://ebookcentral.proquest.com/lib/utranscomm-ebooks/detail.action?docID=456896.
Created from utranscomm-ebooks on 2018-01-30 00:53:03.
Accuracy Assessments 71
an overall mean error rate of less than 0.50% and a systematic error of less than
0.50%, plus a random error of less than ±1.00%, at a 95% confidence level for
interval recordings. (The precise figures given here are for illustration only.)
A specification may sometimes call for the TMS to show “no systematic
bias,” “no bias,” “no long-term cumulative error,” or something similar. As ex-
plained, all TMSs have systematic errors or bias; hence, none of these points is
actually achievable. The purpose of such a requirement is usually to disallow the
calibration of the equipment in such a manner as to establish a permanent bias
in favor of one party or another. There is an alternative way of approaching these
types of requirement through a deliberately biased piece of equipment; this is
discussed further in Section 4.16.
If such a requirement is confirmed, it is customarily interpreted as meaning
“no significant bias,” where “significant” refers to the performance specification.
Dalgleish, Michael, and Neil Hoose. Highway Traffic Monitoring and Data Quality, Artech House, 2008. ProQuest Ebook
Central, http://ebookcentral.proquest.com/lib/utranscomm-ebooks/detail.action?docID=456896.
Created from utranscomm-ebooks on 2018-01-30 00:53:03.
72 Highway Traffic Monitoring and Data Quality
As a working rule, this may be taken to mean that the bias should be less than
some proportion of the stated accuracy requirement, for example, 50% or 33%.
CIM99.7% is contained wholly within the specified accuracy limits. This test fur-
ther extends the even- and two-sigma probability tests by mandating that the
true population mean must lie within the specification at a 99.7% confidence
level (i.e., will be incorrect only 3 times in 1,000). It is a rigorous assessment
of bias since the mean from this machine will be within the specification 997
times out of every 1,000. In other words, it is practically never in error. It can be
remembered by the fact that the three is the z factor of 3.00
Dalgleish, Michael, and Neil Hoose. Highway Traffic Monitoring and Data Quality, Artech House, 2008. ProQuest Ebook
Central, http://ebookcentral.proquest.com/lib/utranscomm-ebooks/detail.action?docID=456896.
Created from utranscomm-ebooks on 2018-01-30 00:53:03.
Accuracy Assessments 73
Zero error
Negative errors Positive errors
Hit accuracy/low precision
and within specification
Figure 4.1 Distribution of errors for a TME with high accuracy and low precision.
Zero error
Figure 4.2 Distribution of errors for a TME with low accuracy and medium precision.
the accuracy test despite having very different accuracy (i.e., the size of the mean
error) and precision (i.e., the range of errors around the mean error).
The position of the mean is irrelevant; all that matters is that the 95% con-
fidence interval lies inside the specification. Figure 4.1 shows the widest distribu-
tion that would pass, albeit using a quite unbiased machine. Figure 4.2 shows
how the distribution can be quite biased and quite wide (i.e., imprecise too) and
yet still pass. The final graph in Figure 4.3 shows a biased but precise machine.
All passes are on the limit.
Dalgleish, Michael, and Neil Hoose. Highway Traffic Monitoring and Data Quality, Artech House, 2008. ProQuest Ebook
Central, http://ebookcentral.proquest.com/lib/utranscomm-ebooks/detail.action?docID=456896.
Created from utranscomm-ebooks on 2018-01-30 00:53:03.
74 Highway Traffic Monitoring and Data Quality
Figure 4.3 Distribution of errors for a TME with low accuracy and high precision.
The restricted mean test reflects the view that a mean well away from the accu-
racy limit will be more robust than a mean close to the limit. The restricted mean
test adds the additional criteria that the mean must be within ±50% of the speci-
fied accuracy limit. The logic of this addition is that the most likely systematic
error is restricted to lying within one-half of the accuracy requirement, which
acknowledges the fact that systematic errors are much more inconvenient than
random errors, whose effect tends to cancel out.
Figure 4.4 shows three notable distributions. The central distribution shows
a wide range for the mean, extending from the lower to the upper accuracy lim-
its, with 95% of the range lying exactly at the specification limit. The two other
distributions show results at each extremity of the range allowed for by the mean.
All distributions between these two will satisfy the restricted mean test.
Dalgleish, Michael, and Neil Hoose. Highway Traffic Monitoring and Data Quality, Artech House, 2008. ProQuest Ebook
Central, http://ebookcentral.proquest.com/lib/utranscomm-ebooks/detail.action?docID=456896.
Created from utranscomm-ebooks on 2018-01-30 00:53:03.
Accuracy Assessments 75
about the same mean error fails the test because the confidence interval of the
mean does not include zero.
Dalgleish, Michael, and Neil Hoose. Highway Traffic Monitoring and Data Quality, Artech House, 2008. ProQuest Ebook
Central, http://ebookcentral.proquest.com/lib/utranscomm-ebooks/detail.action?docID=456896.
Created from utranscomm-ebooks on 2018-01-30 00:53:03.
76 Highway Traffic Monitoring and Data Quality
The main problem is that this test will produce a failure result as sample
size increases beyond a critical point. In fact, the test is the result of a misun-
derstanding of the nature and meaning of the confidence interval of the mean.
Since all machines have a bias or systematic error, zero will eventually lie outside
the confidence interval of the mean of any machine, given a large enough sample
size. Therefore, ultimately all machines can be made to fail this test.
Figure 4.6 illustrates this point. When just a few samples are taken, the
confidence interval is very flat and wide, as shown on the diagram for n = 5
Dalgleish, Michael, and Neil Hoose. Highway Traffic Monitoring and Data Quality, Artech House, 2008. ProQuest Ebook
Central, http://ebookcentral.proquest.com/lib/utranscomm-ebooks/detail.action?docID=456896.
Created from utranscomm-ebooks on 2018-01-30 00:53:03.
Accuracy Assessments 77
samples. At this number of samples, the TMS fails the specification since the
95% confidence interval extends outside the upper limit. After 30 samples, the
estimate of the population mean has not moved much, but the confidence inter-
val has now reduced in width, and the (same) TMS can be stated to be within
specification for accuracy, including both zero-included and restricted mean re-
quirements. Finally, after 100 samples have been taken, again the mean has not
moved much, but the confidence interval is now much narrower and therefore
excludes zero, making the TMS now fail the zero-included test.
All three results apply to the same TMS; all that has changed is the number
of samples in the error survey. This hopefully gives a better understanding of
what the sample number does to the estimate of the confidence interval of the
mean and how the zero-included test is really not reliable. This is further elabo-
rated in the next section.
Consider the example in Table 3.1. The systematic error test is assessed
as –0.61% ± 0.74% (i.e., –1.35% to +0.13%). Therefore, 0.00% is included
in the range. This TMS is thus assessed as passing the zero-included test. Now
assume the supplier makes a large improvement to the TMS random error rate.
Assume that the standard deviation test result is improved to ±0.25%, which is
almost three times as good as the previous ±0.71%. This means that the readings
coming from the machine are more precise or consistent from reading to read-
ing than previously. At the same time, assume that the supplier also manages to
improve the systematic error rate to 0.50%, more than a 20% improvement on
the previous value of –0.61%.
Taking the improvement in precision and reworking the calculation for the
confidence interval, we have
Dalgleish, Michael, and Neil Hoose. Highway Traffic Monitoring and Data Quality, Artech House, 2008. ProQuest Ebook
Central, http://ebookcentral.proquest.com/lib/utranscomm-ebooks/detail.action?docID=456896.
Created from utranscomm-ebooks on 2018-01-30 00:53:03.
78 Highway Traffic Monitoring and Data Quality
SD 0.25
CIM 95 = t 95,n × = ±2.57 × = ±0.26%
n 6
Now, using the revised mean error of –0.50%, the confidence interval of
the mean is assessed as –0.50% ± 0.26% (i.e., –0.50% – 0.26%) and –0.50%
+ 0.26% (i.e., –0.76% to –0.24%). Since the confidence interval of the mean
no longer includes 0.00%, the TMS is now declared as failing the zero-included
test. This is shown graphically in Figure 4.7, where the taller curve represents the
error distribution from the improved TMS.
In other words, by improving the random error rate more than the system-
atic, we caused the TMS to fail the zero-included-in-range test. In the same way,
it can be shown that increasing the random error rate, either alone or in conjunc-
tion with a smaller increase in the systematic error rate, will make passing the
zero-included test more probable.
• Repeat the test using a subset of the data collected to reduce the sample
size. This will in effect increase the divisor in the confidence interval esti-
mation and produce a wider distribution. This will clearly only work if
Copyright © 2008. Artech House. All rights reserved.
Dalgleish, Michael, and Neil Hoose. Highway Traffic Monitoring and Data Quality, Artech House, 2008. ProQuest Ebook
Central, http://ebookcentral.proquest.com/lib/utranscomm-ebooks/detail.action?docID=456896.
Created from utranscomm-ebooks on 2018-01-30 00:53:03.
Accuracy Assessments 79
the mean is 50% or less than of the tolerance or accuracy requirement after
the mean from the reduced sample set is considered.
• The technique implied above in Section 4.14 can be applied to the TMS.
This involves introducing random errors into the raw data (inside or
outside the TMS) and producing adjusted data. This will increase the de-
nominator and again produce a wider distribution. Again, it will only
work if the mean is 50% or less than of the accuracy requirement.
Both of these actions aim at exploiting the weakness in this test. This is
permissible as long as the repeated tests performed on the adjusted data are then
in compliance with the systematic and random error requirements. However, the
client could be asked whether he would rather adopt the restricted mean condi-
tion, which achieves the same goal (i.e., halving bias from the specified limit),
while not having sample size problems. If the restricted mean test is adopted,
there may be implications for the cost of the TMS, as well as its maintenance
and verification.
In any event, the verification team would do well to watch the sample size
and keep it to the minimum necessary to enable a pass at the specification limit,
either positive or negative. As soon as this point has been reached, no more
samples should be taken. Incidentally, the remarks in this section, while focused
on vehicle counting, also apply to most vehicle and traffic stream parameters.
But in the specific case of “shadow tolling,” a problem arises: when the system
turns in systematic overcounts, the shadow toll payer can become dissatisfied
that he is continually paying for vehicles that don’t exist. If the payer is the
taxpayer, then the political aspect of paying for vehicles that don’t exist can be
problematic.
This problem is similar to the measure in any commodity subject to weights
and measures control. The criterion changes from best estimate to guaranteed
goods for a given price. In other words, the specification becomes, say, +0.00%
– 2.00%.
This of course relies on equipment being available to achieve such a goal,
which is not necessarily straightforward for a counting device, although it is pos-
sibly easier for measuring continuous variables. However, the advantage is that
Dalgleish, Michael, and Neil Hoose. Highway Traffic Monitoring and Data Quality, Artech House, 2008. ProQuest Ebook
Central, http://ebookcentral.proquest.com/lib/utranscomm-ebooks/detail.action?docID=456896.
Created from utranscomm-ebooks on 2018-01-30 00:53:03.
80 Highway Traffic Monitoring and Data Quality
the payer will (at, say, a 95% confidence level) never pay for vehicles that weren’t
there or didn’t meet some other variable criterion.
Rounding this up to 22, which is now larger than the 6 in the first survey,
we split the difference and add 10% (i.e., 14 + 1 = 15). Substituting student’s t
Copyright © 2008. Artech House. All rights reserved.
n =t 2
× = 2 .15 2
× = 4.6 × 3.31 = 15.3
0.39
p ,n
CIM p
This figure is close to the assumption of 15, so we accept this result and
round up to the nearest integer (i.e., 16 samples are required). From this may be
deducted the 6 samples already taken, leaving another 10 to be done (i.e., 1 hour,
40 minutes more). At the rate of 1,157 vehicles per hour, this means a minimum
sample size of about 3,085 vehicles.
After the total 16 observations are made, the mean error and standard de-
viation should be recalculated and then reassessed for the minimum samples. For
Dalgleish, Michael, and Neil Hoose. Highway Traffic Monitoring and Data Quality, Artech House, 2008. ProQuest Ebook
Central, http://ebookcentral.proquest.com/lib/utranscomm-ebooks/detail.action?docID=456896.
Created from utranscomm-ebooks on 2018-01-30 00:53:03.
Accuracy Assessments 81
example, let’s assume that after 16 samples we have a result of a mean error of
0.60% and a standard deviation of 0.73%:
2
SD 0.73
2
n =t 2
× = 2.15 ×
2
= 4.6 × 3.33 = 15.3
0.40
p ,n
CIM p
This confirms that the number of samples is safely 16, after taking account
of all the data.
This means just another 375 (1,531 – 1,157) vehicles need be surveyed. As
before, the calculation should be rechecked when that number of observations is
complete and the updated value for mean error is available.
For example, let’s assume the new mean error rate is 0.62%:
Dalgleish, Michael, and Neil Hoose. Highway Traffic Monitoring and Data Quality, Artech House, 2008. ProQuest Ebook
Central, http://ebookcentral.proquest.com/lib/utranscomm-ebooks/detail.action?docID=456896.
Created from utranscomm-ebooks on 2018-01-30 00:53:03.
82 Highway Traffic Monitoring and Data Quality
merate the vehicle count and/or classification against the TMS count and
compile a pass/fail for compliance with the accuracy specification.
2. A speed-measuring TMS is to be audited using a calibrated handheld
radar gun. The process is manual, with each pair of readings, the value
from the TMS, and the accepted reference value from the speed gun
being entered into a spreadsheet program on-site. After the survey is
complete, the resulting mean error and standard deviation are used to
calculate compliance with the specification.
In both cases, the manual element of the work (i.e., the enumeration in
the first case and the data collection in the second case) can be stopped if the
mean and the confidence interval of the mean are calculated after each vehicle is
assessed. In the case of the vehicle counts, this would be one of the automated
Dalgleish, Michael, and Neil Hoose. Highway Traffic Monitoring and Data Quality, Artech House, 2008. ProQuest Ebook
Central, http://ebookcentral.proquest.com/lib/utranscomm-ebooks/detail.action?docID=456896.
Created from utranscomm-ebooks on 2018-01-30 00:53:03.
Accuracy Assessments 83
4.20 Calibration
Calibration is a process whereby a TME or the reported data is adjusted such that
the data provides the best estimate of the true values. Calibration may involve a
process with the TME (e.g., speed calibration) and/or a process in a downstream
computer system (e.g., count adjustment).
The derivation of calibration factors is a nontrivial exercise, especially
where the accuracy of the raw data is good, for example, ±1% or better. Many
hours of manually enumerated (24 hour) data from video-recorded observations
are required in order to determine calibration factors with suitable confidence
Copyright © 2008. Artech House. All rights reserved.
intervals. And this work must be redone if there are significant changes in traffic
flow at the site.
It is possible that the confidence interval for the calibration factors may be
wider than their effect. In this case, a decision not to use the calibration factors
may be made since the application of the factors to the raw data adds insignifi-
cant value at the required confidence level. This will often be the case with very
accurate TMS installations.
Calibration factors have to be reviewed and resurveyed at appropriate time
intervals and involve considerable survey efforts. In general, the more accurate the
raw TMS data is, the greater the cost of providing usable adjustment factors.
Calibration is a different process from assessment or verification. The de-
sign of an assessment process is such that qualitative data arising may be un-
suitable for calibration use. When verification is performed with nonrandom
Dalgleish, Michael, and Neil Hoose. Highway Traffic Monitoring and Data Quality, Artech House, 2008. ProQuest Ebook
Central, http://ebookcentral.proquest.com/lib/utranscomm-ebooks/detail.action?docID=456896.
Created from utranscomm-ebooks on 2018-01-30 00:53:03.
84 Highway Traffic Monitoring and Data Quality
samples, the sample data should not be used for calibration without careful
consideration.
Since LPSEP can usually only be entered to the nearest centimeter, the
value of 453 cm would be used. A small residual systematic error would remain,
and subsequent tests should be expected to yield a continuing variation in the
order of this difference. This is then an example of a systematic error that cannot
be removed.
Copyright © 2008. Artech House. All rights reserved.
Dalgleish, Michael, and Neil Hoose. Highway Traffic Monitoring and Data Quality, Artech House, 2008. ProQuest Ebook
Central, http://ebookcentral.proquest.com/lib/utranscomm-ebooks/detail.action?docID=456896.
Created from utranscomm-ebooks on 2018-01-30 00:53:03.
5
Collecting Data in Groups
5.1 Introduction
Binning is the process that collects data into groups by incrementing a count
(bin) based on a vehicle parameter. For example, a vehicle measured at a speed
of 48 kph would be placed in a speed bin with limits of 40 to 50 kph. Binning is
often used to conserve memory and keep communication charges low. It is most
commonly used for:
Each of these binning counts use the same binning error theory, but the
different distributions of the parameters being measured will require different
detailed analysis.
It is unlikely in a modern machine for binning to be done in error. For
example, it is unlikely that a vehicle reported to be traveling at a speed of 48 kph
will be placed in the speed bin with limits from 51 to 65 kph. But because of the
85
Dalgleish, Michael, and Neil Hoose. Highway Traffic Monitoring and Data Quality, Artech House, 2008. ProQuest Ebook
Central, http://ebookcentral.proquest.com/lib/utranscomm-ebooks/detail.action?docID=456896.
Created from utranscomm-ebooks on 2018-01-30 00:53:03.