Professional Documents
Culture Documents
Evaluation of Technologies For Road Profile Capture Analysis and Evaluation - Chin Et Al. 2015
Evaluation of Technologies For Road Profile Capture Analysis and Evaluation - Chin Et Al. 2015
Abstract: Many transportation agencies in the United States are transitioning to using the International Roughness Index (IRI) in pavement
smoothness evaluation as part of an incentive/disincentive program for contractors paving roads. This study evaluates the use of digital levels,
inertial profilers, inclinometer-based profilers, and terrestrial laser scanning (TLS) in pavement profile capture and analyses. Traditionally, anal-
yses evaluated elevation differences between profiles. However, evolving standards and procedures perform incremental slope-based analyses
to calculate IRI and correlate between profiles, which correspond better to vehicle response. The results of this study show consistent elevation
profiles between all of the devices, when operated properly. However, digital levels cannot be used efficiently to capture the road profile at high
sampling (,10 cm) intervals for long profiles. Thus, TLS shows additional advantages compared with the other techniques, including the ability
to collect a large, dense set of data relatively quickly for the entire roadway and surrounding areas; hence, the data can serve multiple purposes,
including validation of other design parameters such as transverse and longitudinal slope. However, TLS requires substantially more processing
and training. DOI: 10.1061/(ASCE)SU.1943-5428.0000134. © 2014 American Society of Civil Engineers.
Author keywords: Light detection and ranging (LIDAR); Laser scanning; Road profile; Roughness; International Roughness Index (IRI);
Inertial profiler.
Introduction traveling in one direction on the highway. The typical values of the
IRI range from 1:5 to 3 m=km for new pavements and from 2.5 to
Pavement smoothness is often the primary focus of the public per- 5.75 for older pavements (Sayers and Karamihas 1998).
ception of road quality; hence, it is an important consideration in The IRI can be determined using measurements from a variety of
roadway construction acceptance. The Transtec Group (2008) dis- devices, which ultimately measure the differences in elevation along
cussed several benefits of smooth roads, including less maintenance, the profile. The algorithm to compute the IRI contains a moving
lower operation costs, less dynamic loading compared with a rough average filter, a quarter-car filter, and the length of the section
surface, improved structural integrity, increased durability, and im- (Sayers and Karamihas 1998). The following are important con-
proved safety for drivers. siderations when calculating the IRI:
As such, many state DOTs now offer incentive/disincentive pro- • The profile data must first be filtered to eliminate outliers (Sayers
grams for completion of smooth pavement surfaces. The requirements and Karamihas 1998), which can create artificial roughness.
for these incentive/disincentive payouts are based on measured • The moving average filter applies a low-pass filter of 250 mm to
smoothness/roughness indices (Transtec Group 2008) such as the smooth the profile by using the average values of adjacent points
International Roughness Index (IRI), which is the most commonly to emulate the enveloping effect of a tire on a pavement surface.
used index for evaluating road surface smoothness because of its • The IRI algorithm is based on the quarter-car model, which
reproducibility. Ultimately, the end-goal of a smoothness index is to includes one-quarter of the car and the mass supported by one
model a typical vehicle response to variations in the road surface tire; this is sometimes referred to as the Golden Car.
(i.e., roughness). • The IRI considers the length of the section measured, which is
The IRI is calculated in units of inches/miles or meters/ why the IRI is in units of slope (Sayers and Karamihas 1998).
kilometers as the average rectified slope referenced to a standard • The localized roughness is determined separately because rough
quarter-car model traveling at 80 km=h (Dyer and Dyer 2008). The sections will be averaged out if a long length is used in reporting
IRI model can be implemented to evaluate the roughness of both new the IRI. Localized roughness is defined as any 7.62-m segment
and existing pavement sections along a profile. For postconstruction, that contains IRI values disproportionately affecting the overall
quality control profiles are evaluated based on profiles acquired for IRI (AASHTO 2010a).
the left wheel path (LWP) and right wheel path (RWP) of a vehicle • The IRI is sensitive to wavelengths from 1.2 to 30 m (Karamihas
2005).
For interested readers, additional information regarding pavement
1 smoothness evaluation indices, the IRI, and the development of
Civil Engineer, AECOM, 250 Apollo Dr., Chelmsford, MA, 01824
(corresponding author). E-mail: abby.chin@aecom.com
a quarter-car model can be found in Sayers and Karamihas (1998).
2
Assistant Professor, School of Civil and Construction Engineering,
Oregon State Univ., 220 Owen Hall, Corvallis, OR 97331. E-mail: michael
Cross-Correlation
.olsen@oregonstate.edu
Note. This manuscript was submitted on May 21, 2013; approved on
February 12, 2014; published online on May 2, 2014. Discussion period While the IRI is used to evaluate pavement smoothness at the project
open until October 2, 2014; separate discussions must be submitted for validation level, cross-correlation is used to determine the accuracy
individual papers. This paper is part of the Journal of Surveying Engi- and repeatability of profiling instruments. Cross-correlation pro-
neering, © ASCE, ISSN 0733-9453/04014011(13)/$25.00. vides more insight into the agreement between profiles than does the
Comparison Analysis Procedures In this study, 15 different sampling intervals were chosen (0.025–0.3 m
in increments of 0.025, 0.4, 0.5, and 0.6 m). Analyses were then run
Comparisons between the four profiling methods can be drawn by to determine the optimal sampling intervals. Initially, each profile
determining the road roughness (IRI values) and the cross-correlation was compared against the other sampling intervals for each wheel
between the profiles. These results allow a closer examination re- path. The outlying profiles tended to occur with sampling intervals of
garding the use of laser scanning data in the analysis of road rough- 0.25 m or more (Fig. 3), indicating that the larger sampling interval
ness. Transverse, cross-slope values can be calculated from the laser for the filter creates a profile that has been artificially smoothed. The
scanning data and slopes can be compared between the laser scanner profiles from Fig. 3 show elevation differences (typically, less than
and RL. 0.006 m) for a 3-m section for profiles derived from each sampling
The profiles were filtered and analyzed using the freely available interval.
PROVAL software, which can analyze profiles to calculate IRI Following this visual analysis, the IRI values from each sampling
values, determine cross-correlation, and extract wavelength content. interval profile were examined (Fig. 4). It is likely that the 0.025-m
To account for vehicle tire size, a 250-mm filter is recommended to sampling interval did not eliminate all of the scanner noise and the
model the tire envelope. This filter was not applied to the level data surface appeared rougher than it actually was. The trend in the
(the spacing of the level data, 300 mm, is larger than the 250-mm IRI curve flattened out between sampling intervals of 0.075 and
filter), or the IBP (the measurement system has a built-in mechanical 0.25 m for both wheel paths. These values also agree very well with
filter because it uses wheels directly on the surface). However, a the IRI obtained from the inclinometer-based profiler. The sam-
250-mm moving average filter was applied to the inertial profiler pling intervals from 0.075 to 0.3 m for both wheel paths provided
data and TLS data, which can sample finer intervals. profiles with no outlying points and were within, or bordered,
To compare the devices and evaluate the suitability of TLS for the 65% range from the IBP IRI.
smoothness determination, the following analyses were performed: A complete data set for the road section offers the ability to
1. Sampling and filtering requirements for TLS: TLS collects collect multiple profiles, calculate cross slopes, and calculate IRI at
data at fixed angular increments; hence, the sampling on the any interval across the roadway. The IRI values for profiles spaced
ground is not uniform (dense closer to the scanner; more sparse every 0.3 m across the roadway are shown in Fig. 5, where some
farther away from the scanner). Furthermore, multiple scan variability in the IRI values to the left and right of the LWP and
setups (with individual georeferencing errors) are combined significant variability slightly to the left and right of the RWP can be
into a single point cloud. To this end, the data need to be seen, indicating that any deviation from the wheel path by the inertial
filtered [using the Bin ‘N’ Grid (Olsen 2011) procedure pre- profiler can influence the IRI values. This may have a greater effect
viously described] and sampled at a regular interval for profile on the cross-correlation values because these are dependent upon the
analysis. Because of the angular and ranging errors inherent in location of the roughness. The variability observed from obtaining
TLS, points collected from TLS will not perfectly model the multiple profiles along the roadway provides insight into the reasons
surface and will result in noise on the order of 3–5 mm (1–s) for the lower cross-correlation values. It also provides a clearer
for most current laser scanners. While all devices have sys- picture of the actual roughness of the road rather than that observed
tematic noise in one form or another, the high-resolution nature solely in the wheel paths.
of TLS data enables seeing this noise. Hence, because of the
significant amount of measurement redundancy, filtering pro-
IRI Comparisons
cesses may help reduce the noise, which is similar to a least-
squares regression of points to a curve, where sample points Fig. 6 provides a comparison of the IRI values from the various
will be present above and below the modeled surface. In instruments. Except for the April 2012 RL survey, all values fall
contrast, selecting points at discrete intervals within the point within 5% of the IBP values. The IBP consistently has the lowest SD
cloud will not be able to reduce the noise. between runs, indicating highest repeatability. The IRI obtained
2. IRI calculation comparisons: All devices were compared with from the RL tends to be higher than the IRI from the other devices
the IRI values determined for the 161-m length on both wheel for both wheel paths. This is likely a result of the larger sample
paths. In the case of instruments with multiple runs, variances spacing (0.3 m). Because only one RL survey for each wheel path
in the IRI values of those devices were also determined. was completed each day, a SD cannot be computed. The TLS data
3. Cross-correlation between profiles and devices: The collected are based on the average IRI from the 0.05–0.125 m sampling in-
profiles from the instruments were run through the PROVAL terval profiles, which were not obtained from additional passes. The
inertial profilers show a large SD compared with the IBP. This is not There were again accuracy problems with the data from the LWP.
surprising given the difficulty in navigating a straight path at The November 2011 inclinometer test had low cross-correlation
highway speeds and the observed variability across the test site values for both wheel paths. In Table 2, the diagonal elements
(Fig. 5). represent repeatability within all of the runs with that instrument on
the survey date. For the TLS data, this repeatability is based on
Cross-Correlation Analysis comparing the 0.025-, 0.05-, 0.075-, and 0.1-m spacing intervals,
rather than separate TLS surveys. The profile selected for the ref-
As discussed in the “Introduction,” the cross-correlation provides erence profile was the profile that showed the highest correlation
a better analysis of the agreement of two profiles than does the IRI with the other runs for that device and date.
alone. Fig. 7 compares the cross-correlation values obtained by
comparing the TLS with varying sample intervals to the RL, in-
clinometer, and inertial profiler as reference profiles. It is likely that
Elevation Correlation Analysis
the results would agree better if (1) there were less variability across
the road surface, minimizing wandering effects (the horizontal A traditional, elevation correlation analysis was completed to
movement of the profiler off of the path); and (2) the profiles were compare the profile elevations with the findings of Chang et al.
completed within a short time window, reducing seasonal effects and (2006) regarding the use of TLS. As discussed in the “Introduction,”
differences in road degradation. the IRI-based cross-correlation focuses on slope variations rather
The comparison with the RL showed the poorest correlation, than elevation variations because slope variations are more appli-
likely a result of the large spacing (0.3 m) of the data set. The inertial cable to ride quality.
profilers and IBP both generally showed good cross-correlation with Two forms of reporting accuracies are presented in Table 3:
the TLS data, particularly between 0.05 and 0.2 m for both wheel The RMS (∼68% confidence interval) and the accuracy expressed
paths. More scatter was observed for the RWP, likely as a result of at a 95% confidence interval (51:96 3 RMS for 1D normally
the higher variability in roughness near the RWP (Fig. 5). However, distributed data). Because the cross-correlation calculation adjusts
the LWP cross-correlation values were consistently lower. the profiles such that the mean elevation value is zero, the RMS
The cross-correlations were determined using various reference/ and 95% confidence intervals were computed with a mean dif-
comparison profiles, as given in Table 2, where the bold values meet ference profile elevation adjustment. The November 2011 in-
current AASHTO standards (90% accuracy and 92% repeatability). clinometer survey was used in this analysis because there was no
Most of these bold values are from the comparison of the RWP. DMI calibration and closed-loop adjustment for the June 2011
Fig. 5. IRI values obtained for longitudinal profiles spaced every 0.075 m across the roadway, with point spacing at 0.075 m
inclinometer survey, and these are critical for elevation comparisons. (0:073 m=km), which as discussed previously were based on
The inclinometer and level profiles show the best agreement. The incremental slope measurements.
TLS shows slightly better agreement to the inclinometer-based
profiler than the level. These values are reasonable given the typ-
Wavelength Analysis
ical relative accuracies (1–s) of the equipment (TLS, nominal RMS
accuracy of 0.005 m; RL, nominal RMS accuracy of 0.001 m; IBP, The profile runs from each device were compared using the power
nominal RMS accuracy of 0:00025=7:62 m 5 0:005 mm for 161 m, spectral density analysis function in PROVAL. The IRI filters
with a closed loop). were applied to all devices, with the 250-mm filter applied to
A second statistical analysis was completed to compare the the TLS and inertial profiler data. Fig. 8(a) shows a comparison
November 2011 and April 2012 RL surveys. The RMS values of the wavelengths for all devices for LWP deviations below
between the elevation measurements (without a mean adjust- 1:5 m=cycle and above 40 m=cycle. Fig. 8(b) shows a compari-
ment) were also low (#0:001 m), which shows good correlation son of the slope wavelengths for RWP deviations below
with the data despite the large difference observed in IRI values 0:9 m=cycle and above 60 m=cycle. These trends are similar to
the LWP except that the differences are much more pronounced
below 0:9 m=cycle. Also, note that at around 12 m=cycle, the
TLS shows some deviation. However, overall, for the range
at which the IRI calculations are most sensitive (1.2–30 m)
(Karamihas 2005), the devices show good agreement in wave-
length content. Furthermore, given that the profile length was 161
m, wavelengths greater than 30 m would require a longer profile
to accurately capture them.
Based on the Nyquist-Shannon sampling theorem, in order to
reconstruct a wavelength of 1.2 m (the lower bound of the range at
which the IRI calculations are sensitive) (Karamihas 2005), at
a minimum, samples would be required every 0.6 m. However, the
Nyquist-Shannon sampling theorem assumes perfect data; this was
not the case in the data sources used, which contained noise. As such,
the previously determined optimal sampling interval of 0.075–0.25
m necessary for TLS data requires 2–8 times more samples than the
theoretical amount required.
Comparisons of the TLS data only for each sampling interval for
the LWP and RWP are shown in Figs. 8(c and d), respectively. These
comparisons were done without the IRI and 250-mm moving av-
erage filter previously applied. Figs. 8(c and d) illustrate that there is
some variability in wavelength content, depending on the sampling
interval. In particular, the 0.175-m sampling interval shows some
Fig. 7. Cross-correlation comparisons between TLS and other devices significant deviations for the LWP. However, for the shorter and
for the (a) LWP and (b) RWP at varying sampling intervals longer wavelengths, these differences are typically less than those
observed when comparing all devices.
04014011-10
IBP_1106 99.02 (0.16) 92.38 (0.21) 92.23 (0.17) 93.37 (0.36) 83.93 (0.66) 91.42 (0.14) 94.22 (0.17) 96.92 (0.29) 69.39 (0.40) 98.00 (0.40)
IBP_1111 92.42 (0.72) 99.32 (0.19) 79.75 (0.85) 84.90 (0.70) 77.94 (0.64) 84.15 (0.58) 83.40 (0.71) 87.39 (0.69) 62.48 (0.46) 93.87 (0.50)
IBP_1206 92.39 (1.06) 80.04 (1.30) 99.24 (0.28) 92.90 (0.75) 86.58 (0.33) 86.61 (0.99) 91.90 (0.93) 95.37 (0.93) 66.30 (0.63) 90.33 (1.08)
RL_1111 95.82 — 87.81 — 91.18 — — — 83.17 — 89.89 — 92.47 — 94.73 — 67.29 — 93.90 —
RL_1204 87.45 — 81.32 — 89.48 — 83.17 — — — 80.22 — 84.81 — 89.27 — 61.55 — 86.57 —
TLS_1110 92.07 (0.88) 84.67 (0.96) 86.37 (0.88) 87.21 (1.46) 78.34 (1.47) 98.56 (0.10) 94.02 (0.50) 90.85 (0.80) 72.12 (0.84) 89.39 (0.72)
IP1 92.80 (2.29) 81.86 (2.26) 91.99 (1.22) 91.27 (0.94) 81.69 (1.00) 92.91 (0.84) 97.51 (1.46) 92.76 (1.67) 72.26 (0.82) 90.63 (1.53)
IP2 94.22 (2.99) 84.91 (3.14) 93.80 (1.24) 92.15 (2.28) 84.76 (0.85) 90.14 (1.70) 94.98 (1.65) 95.66 (3.40) 70.41 (2.29) 93.26 (2.64)
IP3 69.64 (0.63) 62.33 (1.32) 65.92 (0.59) 64.36 (1.10) 58.19 (0.85) 70.48 (2.43) 71.98 (2.43) 76.81 (1.14) 95.58 (1.41) 69.18 (0.77)
IP4 95.49 (1.81) 92.07 (3.06) 90.61 (2.19) 90.94 (0.87) 82.94 (1.96) 87.47 (1.31) 89.56 (2.02) 92.27 (2.81) 68.20 (1.95) 96.94 (1.64)
Note: Bold values meet current AASHTO (90% accuracy, 92% repeatability) requirements (AASHTO 2010b); RL 5 nominal RMS accuracy of 0.001 m; numbers 1106, 1111, etc., 5 experiment dates: 1106
5 November 6; 1111 5 November 11, and so on.
J. Surv. Eng.
Discussion contrast, a sample interval that is too small will not filter out the
data enough and there will be scanner noise visible in the profile.
This section summarizes the following insights obtained with re- The optimal sampling interval determined from this study is in
spect to the initial motivations and other considerations found as the range of 0.075 to 0.25 m.
a result of this study: • Cross-correlation: The TLS-derived profiles were proven to be
• Filtering: Point filtering such as the Bin ‘N’ Grid (Olsen 2011) consistent with the other techniques based on the high cross-
method described in this study can reduce the noise associated correlation values. In the wavelength analysis, the TLS showed
with TLS. Given the agreement in IRI values and cross- good agreement compared with the other devices.
correlation values with the other systems, it can be concluded • Implications of multiple profiles across the road: The TLS
that TLS filtering can produce data of sufficient accuracy for road showed significant variations in IRI values depending on where
profile evaluations. This agreement is achievable despite the the profiles were obtained across the road. This improved knowl-
higher noise in TLS primarily because the laser footprint is much edge explains the difficulty in achieving cross-correlation with
smaller than the tire envelope filter (250 mm) applied as part of other devices when wandering from the path.
the IRI calculation. However, it should be noted that a time-of- The following additional insights were found in this study re-
flight system, although successful for IRI calculations, will have garding laser characteristics, environmental conditions, surveying
limitations compared with phase-based systems for other pave- considerations, and processing considerations:
• Laser characteristics: Dark surfaces at long ranges are problem-
ment metrics of interest including the detection of small cracks
and microsurface characterization. Nonetheless, a similar Bin ‘N’ atic for some scanners because they do not reflect light well.
Grid (Olsen 2011) filtering scheme could be implemented with Hence, scanning should be performed at close range (,50 m) to
the pavement surface for best results.
phase-based data. • Environmental conditions can influence the results of TLS data
• Sample interval: A sampling interval that is too large will filter
collection. Many of these effects (temperature, pressure, and
out surface texture and create an artificially smooth profile. In
relative humidity) can be corrected for in the instrument or by
Table 3. Statistical Analysis of Profile Elevations for TLS, Level, and only using data for short ranges. Wet pavements will generally
Inclinometer yield poor scanning results, as do conditions where refraction is
present, for example, as a result of steam, precipitation, or heat
TLS to level Inclinometer Inclinometer rising from surfaces. When evaluating road profiles, it is impor-
(mm) to level (mm) to TLS (mm) tant to note that significant changes in roughness can occur with
Parameter Left Right Left Right Left Right weather effects and traffic volume over time.
• Surveying strategies: Scans should be spaced close together to
RMS 5.6 5.0 2.9 2.3 4.5 5.0
minimize oblique scanning on the road surface. Hence, more
Confidence (95%) 11.0 9.9 5.6 4.6 8.9 9.9
lower-resolution scans are better than fewer higher-resolution
Fig. 8. Slope wavelength power spectral density comparison for (a) LWP (for all devices); (b) RWP (for all devices); (c) LWP (for TLS sampling
intervals); (d) RWP (for TLS sampling intervals)