GPS Solut (2018) 22:7

https://doi.org/10.1007/s10291-017-0667-9

ORIGINAL ARTICLE

Analysis and modeling GPS NLOS effect in highly urbanized area

Li‑Ta Hsu1   

Received: 10 January 2017 / Accepted: 13 October 2017 / Published online: 4 November 2017
© Springer-Verlag GmbH Germany 2017

Abstract  Current GPS positioning accuracy in urban Introduction

areas is still unsatisfactory for various applications, includ-
ing pedestrian navigation and autonomous driving. Due The booming of transportation robotics such as unmanned
to the ineffectiveness of special corrector designs against aerial vehicle (UAV) and driverless car pushes the require-
non-line-of-sight (NLOS) reception, the research regarding ment of accuracy and precision of low-cost global navigation
NLOS signals has been increasing in the recent years. This satellite system (GNSS) receivers. Multipath and non-line-
study first develops an algorithm to detect NLOS signals of-sight (NLOS) receptions still are the main challenges for
from the pseudorange measurements by using a 3D building GNSS receiver in urban canyon (Groves 2013). Especially in
model, ray-tracing simulation, and known receiver position. Asian super urbanized cities such as Hong Kong and Tokyo,
According to the analysis of 24 h of collected NLOS data, the challenges from the signal reflection at the dense sky-
a new finding is that NLOS pseudorange delay is highly scrapers further increasing the difficulty. Figure 1 shows the
correlated with the elevation angle of satellite instead of positioning data calculated by a commercial GNSS receiver,
the received signal strength. Thus, we further propose an u-blox M8 receiver (GPS/Beidou mode), and collected in
innovative NLOS model using two variables, the elevation Kowloon, HK. Comparing the top and bottom panels of the
angle and the distance between the receiver and building that figure, the positioning accuracy in urban area is much worse
reflect the NLOS. The proposed model is evaluated in both than that in the open area. It is evident the poor GNSS posi-
pseudorange and position domains. Based on the experiment tioning result is caused by the GNSS signal blockage and
results regarding pseudorange error, the difference between reflection from the surrounding buildings.
the proposed model and the collected NLOS measurement Multipath and NLOS are different effects (Groves 2013),
is very small. Finally, the proposed model is applied to a which are shown in Fig. 2. Multipath contains both the direct
hypothesis-based positioning method and achieves about and reflected signals, while NLOS contains only the latter
6.3 m in terms of horizontal positioning accuracy, which is one. Even for a low-cost GNSS receiver, the multipath effect
only slightly worse than the method applied with ray-tracing can be mitigated by sophisticated receiver correlator designs
simulation. (Veitsel et al. 1998; Zhdanov et al. 2002). The principle of
the correlator design is comparing the early, prompt and late
Keywords  NLOS · Multipath · Urban area · Ray-tracing · channels in code tracking loop. In the other words, it com-
Localization pares the direct signal with reflected one. Unfortunately, this
design does not mitigate the NLOS effect at all because the
NLOS contains only the reflected signal. Thus, the research
focused on detection and mitigation of NLOS is increas-
* Li‑Ta Hsu ing. Detecting the NLOS using carrier to noise ratio (C/N0)
lt.hsu@polyu.edu.hk measurements of dual-polarization antenna is discussed in
1
Interdisciplinary Division of Aeronautical and Aviation
Jiang and Groves (2014). An advance tracking algorithm,
Engineering, The Hong Kong Polytechnic University, i.e. vector tracking, has been proposed to detect and mitigate
Kowloon, Hong Kong NLOS effect in signal processing stage (Benson 2007; Hsu

13
Vol.:(0123456789)
7 
Page 2 of 12 GPS Solut (2018) 22:7

dynamic test in urban area Due to the rise of smart cities, the 3D city models are
rapidly developed and became widely available. Recent
research dealing with multipath and NLOS utilizes 3D map-
ping and is called as 3D mapping aided (3DMA) positioning
methods. Comprehensive related work on 3DMA position-
ing method can be found at Groves et al. (2015) and Breßler
50 meters of error in 2D et al. (2016). One of the most well-known 3DMA methods
is shadow matching proposed by university college London
(UCL). It takes advantage of the 3D city model to generate
Ground truth building boundary in the Skyplot to help predicting satellite
GNSS result visibility (Groves 2011; Wang et al. 2012, 2013, 2015). The
improvements of shadow matching in land, aviation, and
static test in open area (shore side) mapping applications are deeply investigated by American
and Israeli researchers (Yozevitch et al. 2014; Isaacs et al.
2014; Yozevitch and Ben-Moshe 2015). Researchers from
5 meters of error in 2D the French institute of science and technology for trans-
port (IFSTTAR) and the national higher French institute of
aeronautics and space (ISAE-SUPAERO) also focused on
improved GNSS positioning accuracy using enhanced 3D
digital map (Ahmad et al. 2013; Betaille et al. 2013; Pey-
Ground truth raud et al. 2013; Peyret et al. 2014). Instead of mitigating
GNSS result or excluding NLOS effect, in the past 3 years, the potential
of using NLOS signal in constructive senses has been pro-
Fig. 1  Performance of a low-cost GNSS receiver obtained dynami- posed. Researchers from Japan and Canada propose to com-
cally and statically from a quadcopter in the urban and open area of bine ray-tracing simulation with hypothesis-based position-
Kowloon, HK, respectively, on September 2016 ing method to further improve the positioning accuracy (Hsu
et al. 2016b; Kumar and Petovello 2014; Miura et al. 2015;
Suzuki and Kubo 2013). The range-based 3DMA uses a ray-
tracing technique to estimate the reflection route of NLOS
Multipath NLOS
signal. The route is then used to correct the NLOS delay
from the biased pseudorange measurement and it further
improves positioning accuracy to about 5 m for pedestrian
applications (Hsu et al. 2016a). However, the range-based
method cannot be easily adapted to low-cost receiver due to
(1) the heavy computational load caused by ray-tracing and
(2) the inaccessibility to 3D building models in real time.
The novelty and contribution of this research is to propose a
NLOS model that can be used for hypothesis-based position-
ing without using ray-tracing in the low-cost devices.
To achieve the goal, the pseudorange error and its C/N0
Fig. 2  Illustration of signal reception of multipath and NLOS effects has to be investigated. In the stage of analysis, the NLOS is
(Hsu et al. 2016b) identified by a given ground truth of the receiver position and
by 3D building models. Namely, it is identified if the ray-trac-
ing simulation indicates that the building model obstructs the
et al. 2015; Kanwal et al. 2010). Consistency check between direct signal transmission path between satellite and receiver,
pseudorange measurements could detect and exclude large and the receiver still receives it. To obtain the NLOS delay
multipath and NLOS effects when the number of clean in the pseudorange domain, all the other errors are elimi-
measurement is sufficient (Blanch et al. 2015; Groves and nated using differential GPS (DGPS) correction. 24 h of GPS
Jiang 2013; Iwase et al. 2013; Hsu et al. 2017). Another raw data at an highly urbanized area is collected to retrieve
direction of study is to add other sensors to compensate the the NLOS signals at different C/N0 from different elevation
inaccurate GNSS positioning result caused by NLOS effect angle. The result reveals an interesting finding. The pseudor-
(Chiang et al. 2013; Gu et al. 2015; Han et al. 2015; Li et al. ange error caused by NLOS is highly correlated to elevation
2015; Sun et al. 2013; Wang and Gao 2010). angle instead of C/N0. This finding provides an inside view

13
GPS Solut (2018) 22:7 Page 3 of 12  7

for modeling NLOS delay as a function of elevation angle between the reference and rover stations is within 100 km.
and distance from the receiver to the building that reflected This research uses HKSC station of SatRef, which is a net-
the NLOS signal. Note that the distance can be roughly given work of GNSS reference stations established by the land
according to different applications. department of HK government as shown in Fig. 3. The dis-
The detection of NLOS signal and its delay in pseudor- tance between each reference station is less than 20 km. By
ange domain is given first. Analysis of 24-h NLOS data is the use of its archived RINEX 3.02 data that are free to HK
conducted next, followed by a description of the innova- publics, the differential correction can be easily generated.
tion NLOS pseudorange error model proposed here, and the
experimental setup and results. Finally, the conclusions and
future work are summarized. Receiver clock bias and thermal noise

After applying the DGPS correction to pseudorange measure-

Estimation of NLOS delays in the pseudorange ment, its receiver clock bias and thermal noise are still required
domain to be corrected. The general LSE uses multiple (at least 4)
measurements to estimate four unknowns including horizontal,
NLOS pseudorange measurement deteriorate by delays vertical position and receiver clock bias as per equation:
including tropospheric errors, ionospheric errors, satellite
clock/orbit bias, receiver clock bias, and our target NLOS. ⎡ x̂ ⎤
⎢ ŷ ⎥
Other errors must be eliminated before the NLOS analysis. 𝐱̂ =⎢ ⎥ = (𝐆T 𝐆)−1 𝐆T 𝝆̃ (1)
ẑ
DGPS and least square estimation (LSE) are used to deal ⎢ ⎥
⎣ b̂ ⎦
with former and latter part of the errors, respectively.

Differential GPS correction where x̂ , ŷ , ẑ represent the estimated receiver position in

latitude, longitude and altitude directions and the b̂ is the
DGPS is a mature technique (Misra and Enge 2011). A refer- receiver clock bias in meters. G is the geometry matrix,
ence station is installed to receive the GPS measurements. containing the unit line-of-sight (LOS) vectors from satel-
The position of the reference station is precisely surveyed. lites to receiver. 𝝆̃ denotes the measured pseudoranges cor-
After generating DGPS correction, the correction is trans- rected by DGPS correction. An approach to better estimate
mitted to rover. The rover then applies it to the common the receiver clock bias is to reduce the unknown x̂ , ŷ , ẑ .
satellites in view to enhance its own positioning accuracy. The receiver is set statically in the experiment so that the
The differential correction can be generated by calculating ground truth of receiver position can be easily determined.
the difference between the raw pseudorange and true (geo- The ground truth is given by the topographic map available
metric) range for a satellite. Multipath and NLOS effects from land department of HK government. The resolution
to the reference station are usually negligible because the of the map is 20 cm. Each point is given with accurate 2D
open-sky environment and the geodetic antenna with choke- coordinate. The height is given by the topography height
ring design as shown in the right panel of Fig. 3. Theo- obtained from Google plus the height of equipment. Thus,
retically, DGPS correction is capable of eliminating satellite the only unknown left in the LSE is the receiver clock bias
clock/orbit, tropospheric, ionospheric errors if the distance as per equation:

Fig. 3  Distribution of SatRef
stations around HK and antenna SatRef HKSC
and environment of HKSC
station

Experiment area

13
7 
Page 4 of 12 GPS Solut (2018) 22:7

if a measurement has C/N0 larger than 40 dB-Hz and it is

b̂ = (𝐆T 𝐆)−1 𝐆T 𝝆̃ (2) identified as LOS by ray-tracing, the measurement will be
However, if the multipath and NLOS biased measurements used to calculate the receiver clock bias. The receiver ther-
are used, the estimated receiver clock bias will be contami- mal noise is neglected because it is much smaller than the
nated. A clean, namely LOS, measurement identification effect of NLOS. Finally, the NLOS delay in pseudorange
is therefore required. Two criteria are used to identify the domain, Δ𝜌NLOS, can be calculated as,
measurement type, and they are the C/N0 ratio and the signal
( )
type determined by ray-tracing simulation. In general, the Δ𝜌NLOS = 𝜌 − Rrcv + 𝜌Corr + b̂ , ∀𝜌 ∈ NLOS (4)
C/N0 of LOS is larger than that of multipath and NLOS. A
40 dB-Hz threshold of C/N0 is set empirically. The possible where 𝜌 denotes NLOS pseudorange measurement, Rrcv
signal transmission type can be categorized as denotes line of sight distance between the satellite and
receiver position and 𝜌Corr denotes the DGPS correction.
ray ∈ {LOS, NLOS} (3) The analysis of the Δ𝜌NLOS is detailed in the next section.
by the use of ray-tracing. The algorithm of rough LOS iden- The detection of NLOS is simply based on the result
tification through ray-tracing is as follows: of ray-tracing simulation from the given ground truth of

Algorithm 1: LOS identification for a measurement using ray-tracing

simulation

STEP1: Prepare a line segment connecting the receiver and the

satellite of the measurement (i).

STEP2: initialize signal type of measurement (i) as LOS, ()

=

STEP3: for all the building model do

STEP3: for all planes (walls) of a building do

STEP4: if the intersection between the line segment and plane

exists then

STEP5: the measurement identified as NLOS, ()

=

break

end if

STEP6: end for planes

STEP7: end for buildings

13
GPS Solut (2018) 22:7 Page 5 of 12  7

pr. error (m)

std of pr. error (m)
Fig. 6  Mean and standard deviation of NLOS pseudorange error
with respect to carrier to noise ratio

mean

Fig. 4  Skyplots with surrounding building information of data col-

lected in Kowloon. In the top panel, the color of satellite trajectory
denotes C/N0, the redder the color the higher is the C/N0 received. In
bottom panel green, red and yellow indicate LOS, NLOS and mul-
tipath signal, respectively

Fig. 7  NLOS pseudorange error with respect to elevation angle of

the 24-h data. The color denotes the carrier to noise ratio of each
point

receiver position. The NLOS is identified if the LOS path

between the satellite and receiver is blocked, but the receiver
still receives it.

NLOS data analysis

Low-cost u-blox M8 receivers are used to record pseudor-

ange measurements. A 3D building model with level of
detail (LOD) 1 used in this research is obtained from land
Fig. 5  Histogram of NLOS delays in pseudorange domain. The total
number of NLOS signal received is 128,054
department of the HK government. Two data are collected
at different locations of urban area of Kowloon, HK. The

13
7 
Page 6 of 12 GPS Solut (2018) 22:7

first and second data consists of 24 h and 30 min of GPS model (Hartinger and Brunner 1999), the signal strength is
pseudorange measurements, respectively. Figure 4 shows a hint to indicate the quality of pseudorange measurement.
the skyplot of data 1 and 2. As shown in the figure, many Figure 6 shows the relationship between NLOS delay with
signals can be received even if its LOS transmission path respect to C/N0. The mean of NLOS error slowly decreases
is obstructed. In the following analysis, only the identified as the C/N0 increases. Its standard deviation follows the
NLOS signal is evaluated. same trend. This figure roughly verifies the reason that using
the sigma-𝜀 model in weighted LSE can improve the GPS
NLOS data 1: case of 24 h positioning result in many cases even in urban canyon.
Error sources, such as atmospheric and multipath effects,
There are 128,054 NLOS measurements detected in the 24-h are correlated with elevation angle. In general, they mono-
data from different satellites as demonstrated in the top panel tonically increase as elevation angle decreases as described
of Fig. 4. Figure 5 shows the histogram of NLOS delays in in the conventional elevation angle-based weighting model
the pseudorange domain. (Special-Committee-159 2001). Figure 7 shows the rela-
The probability density distribution (PDF) of Fig.  5 tionship between NLOS delay with elevation angle. It is
exhibits the probability of an observed NLOS resulting in clear that the NLOS decreases as the elevation increases,
a corresponding pseudorange error. As shown in the figure, in the other words, using the conventional elevation model
more than 70% of the NLOS delay is less than 50 m. Note in weighted LSE is also able to mitigate the NLOS effect.
that this PDF cannot show the probability of an unknown As shown in the figure, when the elevation angle is between
measurement with a certain pseudorange error to be a NLOS 10° and 20°, its C/N0 is usually lower than 30 dB-Hz. This
measurement. It is evident that NLOS delay cannot be mod- phenomenon proves the agreement that both applying C/N0
eled as Gaussian distribution. By the use of maximum likeli- or elevation angle-based weighting models can improve GPS
hood function, the probability density distribution of NLOS positioning when NLOS signal is present. However, it is not
delay can be described as a Gamma distribution: necessary that the lower the elevation angle is, the lower the
carrier to noise ratio is. Figure 8 shows the different behav-
1 x
f (x|k = 0.81, 𝜃 = 49.23) = x k−1 − 𝜃
e , (5) iors between the received signal strength of LOS and NLOS
𝛤 (k)𝜃 k
using the same commercial GNSS receiver. As shown in the
where 𝛤 (⋅) denotes the Gamma function. Its non-Gaussian figure, the mean of C/N0 of LOS increases as the elevation
property shows the difficulty of mitigating NLOS effect. increases. Its standard deviation decreases as the elevation
As suggested in the popular weighting model, the sigma-𝜀 increases. In the other words, the quality of LOS is more
stable at higher elevation angle. The NLOS data shows a
very different behavior. As the elevation increases, the C/N0
LOS does not increase following the behavior shown in the data
of LOS. The standard deviation of NLOS C/N0 also does
not show the obvious tendency with elevation variation. The
reason is that the NLOS is a reflected signal. The material of
mean of C/N0
the reflecting surface can greatly change the signal strength
mean of C/N0 ± 1σ of reflected signals. Comparing the top and bottom panels,
points collected
one might conclude that C/N0 could be used to distinguish
the LOS or NLOS signal as the elevation angle increase.
This conclusion is similar to the finding in Yozevitch et al.
NLOS (2016) on LOS/NLOS classification using machine learning
approach.
The statistic of a large amount of data can reveal the gen-
eral characteristic of NLOS delay. It will be interesting to
study whether short data would comply with the summarized
characteristics or not.

NLOS data 2: case of 30 min

In the 30-min data, there are 1648 NLOS measurements

detected from three different satellites as shown in the bot-
Fig. 8  Comparison of LOS and NLOS in terms of C/N0 with regards tom panel of Fig. 4. There are 4 LOS, 3 NLOS and 1 mul-
to elevation tipath signal in this case. Similar analysis is conducted to the

13
GPS Solut (2018) 22:7 Page 7 of 12  7

Fig. 9  Three NLOS signals

from different satellite are
collected. Left panel denotes
histogram of NLOS delay and
right panel denotes mean and
standard deviation of NLOS
pseudorange error with respect
to carrier to noise ratio of the
30-min data

Wall of
building GPS signal

normal of wall

Receiver GPS signal

Ground

Fig. 11  Illustration of a reflecting signal that followed the law of

reflection

NLOS is different from multipath. For NLOS, only the delayed

Fig. 10  NLOS pseudorange error with respect to elevation angle of (reflected) signals comes into the receiver, which could eas-
the 30-min data. The color denotes the carrier to noise ratio of each ily deceive the receiver that the NLOS signal is LOS signal.
point
According to the algorithm of estimating C/N0 (Sharawi et al.
2007), the carrier strength is calculated by accumulating the
30-min data as shown in Figs. 9 and 10. To observe Fig. 9, power of I and Q channels. If NLOS signal are continuously
the NLOS pseudorange delay does not decrease as the C/N0 present in the environment, its C/N0 could also be strong. Thus,
increases. Instead, it remains a similar delay with different the strong C/N0 is not necessary indicating a clean measure-
C/N0. Interestingly, the pseudorange delay is still correlated ment. It could still contain tens of meters of pseudorange error.
with the elevation angle as shown in Fig. 10, namely the higher The previously obtained characteristic from the 2-h data of
elevation angle, the smaller the NLOS pseudorange. Based NLOS delay varying with C/N0 is due to the C/N0 being cor-
on the principle of the GPS signal tracking loop, the effect of related with the elevation angle. In fact, C/N0 could increase

13
7 
Page 8 of 12 GPS Solut (2018) 22:7

as the elevation angle increased because of the longer signal satellite and receiver is large compared to the distance
transmission path and the pattern of antenna gain. As a result, between the LOS and reflecting signals, the blue line is not
it can be concluded that elevation angle is a dominant factor only parallel to the greens but also same length. Therefore,
to NLOS pseudorange errors. the NLOS delay (𝛾 ) in the pseudorange domain is the red
lines:

Modeling of NLOS 𝛾 = 𝛾1 + 𝛾2 (6)

where 𝛾1 and 𝛾2 are:
This section tries to derive a model to represent NLOS delay
using the dominant factor, i.e. the elevation angle, previously
𝛾1 = 𝛼 sec 𝜃0 (7)
verified. In highly urbanized areas, most of the buildings use 𝛾2 = 𝛾1 cos(𝜃1 + 𝜃2 ) (8)
excessive glass windows as a face of modern architecture.
where 𝛼 denotes the distance between the receiver to the
We, therefore, assume the GPS signal reflection follows the
obstacles (buildings) that reflected the navigation signal.
law of reflection to model the NLOS signal. An illustration
Angle of reflection 𝜃1 equals to the angle of incidence 𝜃2
of a perfect reflecting signal received by a receiver is given
because the law of reflection is assumed. In addition, since
in Fig. 11.
the green and blue lines are parallel, 𝜃1 and 𝜃0 are alternate
The blue line represents a LOS signal. Green and red
interior angles, namely 𝜃1 = 𝜃0. Assuming the wall of build-
lines are the reflecting signal. Due to the distance between
ing is perpendicular to the ground, thus the normal of wall is
parallel to ground resulting 𝜃2 = 𝜃ele. 𝜃ele denotes the eleva-
tion angle of the satellite. Finally, 𝜃ele = 𝜃2 = 𝜃1 = 𝜃0; thus,
the NLOS delay can be represented as:
( )
𝛾 = 𝛼 sec 𝜃ele 1 + cos 2𝜃ele (9)
In (9), only the distance between the receiver to the obstacle
is unknown to a GPS receiver. Fortunately, this distance 𝛼
could be roughly obtained from the 2D maps that are widely
available to vehicle navigator and pedestrian with smart-
phone. According to road planning standards and guidelines
of HK government, the general road width for primary and
district distributor is between 6.75 and 13.5 m (HKPSG
2016). The path width of pedestrian walk is between 3.5
and 4.5 m (HKPSG 2016). Thus, a proper value of 𝛼 in HK
will be from 3.5 to 18 m for vehicle applications. For the
autonomous driving, the LiDAR sensor will provide the dis-
tances between vehicle and buildings, which can be used by
Fig. 12  Pseudorange error caused in the proposed NLOS model if the proposed NLOS model.
alpha is not accurate

Fig. 13  NLOS can be fre-

quently observed in typical
urban residential area in Asian
urbanized city. The left panel
denotes the environment that
the data were collected, and
right panel indicates the install-
ment of the path antenna

13
GPS Solut (2018) 22:7 Page 9 of 12  7

NLOS of SV01
Wall 4

Wall 3 NLOS of SV11

Wall 2

22m
Wall 1

Receiver position

theoretical value from the NLOS model

mean of collected NLOS delay
Wall 1 collected NLOS delay
End time

Wall 2

Wall 3 Wall 4
Start time

Fig. 14  Example of a received NLOS signal. Top panel demonstrates

signal traveling paths reflected at four different walls by ray-tracing
simulation, and bottom panel shows the NLOS delay of SV 27 that Fig. 15  Example of the two received NLOS signals reflected at same
travelled from 0° to 85° of elevation angle building. Top panel shows the ray-tracing simulation of NLOS from
SV01 and SV11 reflected at Wall 1. Bottom panel shows the collected
NLOS delay and theoretical value calculated by the proposed model
Considering an example, the alpha will be set to 18.2 m
based on the map information, but alpha measured by
LiDAR is 15.6 m. In the case of elevation angle of 61.6° the proposed NLOS model in pseudorange and position
and the alpha discrepancy of 2.6 m, the error introduced domains, respectively.
to the proposed NLOS model is 2.5 m. To further dis-
cuss other cases, a simulation of error introduced under Verification of the NLOS model in pseudorange domain
different elevation angle and alpha discrepancy is shown
in Fig. 12. As can be seen, the newly introduced error is In order to evaluate the proposed model in different elevation
smaller the higher of elevation angle and smaller the alpha angle, a long-time data with NLOS signal must be collected.
discrepancy. Figure 13 demonstrates the environment that the data were
collected. The antenna is attached with a stick and put out-
side of the window for long-time data collection.
Experiment result Figure 14 shows an example that the NLOS signal from
a same satellite that travelled from 0° to 85° of elevation
A commercial GPS receiver, u-blox M8, is deployed to angle.
collect NLOS data. Two different locations at a highly The x-axis of the bottom panel of Fig. 14 is the eleva-
urbanized city are selected to evaluate the performance of tion angle of measurement, and the y-axis is the NLOS

13
7 
Page 10 of 12 GPS Solut (2018) 22:7

65m

42m

39m

26m

13m

0m
Fig. 16  Mean and standard deviation of all the NLOS delays
reflected at Wall 1 from different SVs 0m 13m 26m 39m 42m 65m

Fig. 18  Result of hypothesis-based positioning method using 3D

building mode, the proposed NLOS model and no aiding information

Table 1  Positioning performance of hypothesis-based positioning

method using ray-tracing, the proposed NLOS model and no aiding
information

Methods Mean (m) SD (m)

Ray-tracing 5.05 2.99

Proposed NLOS model 6.27 3.31
No aiding information 8.67 3.93

the proposed model. The distances between the receiver and

the walls (𝛼) are about 22, 29.5, 13.1 and 10.5 m for Wall
1 to 4, respectively. As can be seen from the bottom panel,
the theoretical values are similar to the mean of collected
values (with a region of 10 m), expecting reflection from
Wall 2. In fact, during the elevation angle from 20° to 30°,
the receiver has a great possibility to receive multiple NLOS
Fig. 17  Ray-tracing simulation of the static experiment signals simultaneously. This effect is called multipathing
NLOS, which is complicated to model due to the undisclosed
delay in pseudorange domain calculated by (5). The black tracking correlator design of commercial receiver. To bet-
dashed line denotes the mean of the collected NLOS delay ter verify the proposed model, signals reflected at the same
of different elevation angle (with a resolution of 10°). The wall but from different satellites are preferred. Most of the
trend of the collected NLOS delay decreases as the eleva- reflections of the long-time collected data come from Wall
tion increases. It is clear that the signal transmission path of 1 shown in the top panel.
SV 27 was not blocked during the elevation angle between Figure 15 shows the cases of the other two satellites, SV01
45° and 75° as shown in the bottom panel. The color of and SV11, that only reflected at Wall 1. The black solid line
each point denotes the time of recording data, the bluer the of bottom panel denotes the NLOS delay calculated from the
point is earlier that signal is collected. The NLOS reflections proposed model. Even through the data is noisy, the mean of
of SV 27 are caused by four different buildings as shown NLOS delay is still close to the proposed model. This result
in the top panel. Thus, the NLOS delays can be separated verifies that if the NLOS is a single refection from a wall, the
into four different groups. The red solid lines of the bottom proposed model can predict the NLOS delay. All the NLOS
panel denote the theoretical NLOS delay calculated from reflected at Wall 1 from different SVs is organized in Fig. 16.

13
GPS Solut (2018) 22:7 Page 11 of 12  7

The theoretical value calculated from the proposed NLOS based on C/N0 could be incorrect; (3) NLOS delay is highly
model, namely (9), is very close to the mean of the data. In the correlated to the elevation angle. The second contribution is
other words, the proposed method can obtain the mean of NLOS to propose an innovative NLOS model. The NLOS model
but not the standard deviation. Interestingly, the standard devia- is based on the elevation angle and the distance between the
tion of the data decreases as the elevation angle increases because receiver and the building that reflected the NLOS signal.
the lower the elevation angle is, the lower the C/N0 is obtained. If the distance can be appropriately selected, the proposed
model can accurately describe the NLOS delay of the meas-
Positioning method using hypothesis‑based positioning urements. The effectiveness of the proposed model is dem-
method onstrated by applying it into a hypothesis-based positioning
algorithm. The experiment result shows the positioning error
A static experiment is conducted in the environment shown only to increase about 1.3 m after replacing the NLOS delay,
in Fig. 17. Many glassy buildings are surrounded to the estimated by ray-tracing, into the proposed model in the par-
receiver. A measurement with NLOS reflection is detected ticle filter-based positioning algorithm.
and simulated in the figure. The distance between the build- However, this result still applies the ray-tracing to identify
ing and receiver is about 36.5 m. whether the measurement is LOS or NLOS. Concerning the practi-
To apply the proposed NLOS model, the position of cal use of the proposed NLOS model, it requires working with other
receiver is essential. The previous work proposes a position NLOS detection algorithms. To reduce the computational load, we
hypothesis-based positioning method (Hsu et al. 2016b). This provide three suggestions to integrate with the proposed NLOS
approach distributes several position candidates and gives model. (1) Using additional memory space: an ideal approach is to
them proper weighting by evaluating their pseudorange simi- integrate the proposed NLOS model with shadow matching posi-
larities between the ray-tracing results and measurements. tioning algorithm (Groves 2011), which only applies the building
This positioning method is used to evaluate the NLOS model boundary (using only memory). (2) Benefiting from Big Data era:
proposed. The NLOS pseudorange delay based on ray-tracing another promising and innovative approach is to implement LOS/
is replaced as the NLOS model. The rest of algorithm is the NLOS classifier trained by unsupervised/supervised machine learn-
same. The positioning results of the hypothesis method using ing algorithm before applying the proposed NLOS model. As sug-
ray-tracing, the NLOS and without any aiding information is gested in Yozevitch et al. (2016), their decision tree-based robust
demonstrated and evaluated in Fig. 18 and Table 1. Only the classifier achieves over 85% of accuracy of NLOS detection. (3)
data with the identified NLOS measurement is evaluated. It Increasing the instrument cost: alternatively, a dual-polarization
is important to note that NLOS identification is required in antenna can be implemented in the localization system to effective
the hypothesis-based positioning method. The three meth- detect NLOS reception (Jiang and Groves 2014).
ods evaluated above used 3D building models to determine An interesting future work could be modeling the residue
the signal type of each measurements. It is evident that the of multipath effect after applying sophisticated correlator
method with ray-tracing simulation certainly obtained the design. The nature of multipath effect is very different with
best performance among the three methods, which is about NLOS reception. To study the multipath residue, a software
5 m of horizontal positioning accuracy. However, its com- receiver is required to take the advantage of its flexibility to
putational load is also much higher than that of the other alter the correlator design of code tracking loop.
two. The hypothesis-based method using the proposed NLOS
model achieves 6.27 m of accuracy, which is slightly worse Acknowledgements  Funding was provided by Hong Kong Polytech-
nic University (HK) (Grant No. 1-ZVKZ).
than the method with ray-tracing. If no any aiding informa-
tion used to correct the NLOS measurements, the positioning
accuracy is about 8.67 m. This result verifies the proposed
NLOS model is capable of replacing the use of ray-tracing in
References
calculating NLOS delay in the pseudorange domain. Ahmad KAB, Sahmoudi M, Macabiau C, Bourdeau A, Moura G (2013)
Reliable GNSS positioning in mixed LOS/NLOS environments
using a 3D model. In: European Navigation Conference (ENC),
Conclusions and future work Vienne, Austria, Apr 2013, pp 1–9
Benson D (2007) Interference benefits of a vector delay lock loop (VDLL)
GPS receiver. In: Proceedings of ION AM 2007, Institute of Navi-
Multipath interference and NLOS reception are major error gation, Cambridge, Massachusetts, USA, 23–25 Apr, pp 749–756
sources when using GNSS positioning in urban environ- Betaille D, Peyret F, Ortiz M, Miquel S, Fontenay L (2013) A new
ments. We first analyze a long-time GPS NLOS data and modeling based on urban trenches to improve GNSS positioning
quality of service in cities. IEEE Intell Transp Syst Mag 5:59–70
summarize as follows (1) the NLOS delay is not fitted to Blanch J, Walter T, Enge P (2015) Fast multiple fault exclusion with
Gaussian distribution. Instead, it is more similar to a gamma a large number of measurements. In: Proceedings of ION ITM
distribution; (2) mitigating NLOS effects by given weighting 2015, Dana Point, California, USA, 26–28 Jan 2015, pp 696–701

13
7 
Page 12 of 12 GPS Solut (2018) 22:7

Breßler J, Reisdorf P, Obst M, Wanielik G (2016) GNSS positioning Peyraud S, Bétaille D, Renault S, Ortiz M, Mougel F, Meizel D, Peyret
in non-line-of-sight context-a survey. In: 2016 IEEE 19th inter- F (2013) About non-line-of-sight satellite detection and exclusion
national conference on intelligent transportation systems (ITSC), in a 3D map-aided localization algorithm. Sensors 13:829–847
1–4 Nov 2016, pp 1147–1154 Peyret F, Bétaille D, Carolina P, Toledo-Moreo R, Gómez-Skarmeta AF,
Chiang K-W, Duong T, Liao J-K (2013) The performance analysis of Ortiz M (2014) GNSS autonomous localization: NLOS satellite
a real-time integrated INS/GPS vehicle navigation system with detection based on 3-D maps. IEEE Robot Autom Mag 21:57–63
abnormal GPS measurement elimination. Sensors 13:10599–10622 Sharawi MS, Akos DM, Aloi DN (2007) GPS C/N0 estimation in the
Groves PD (2011) Shadow matching: a new GNSS positioning presence of interference and limited quantization levels. IEEE
technique for urban canyons. J Navig 64:417–430. https://doi. Trans Aerosp Electron Syst 43:227–238
org/10.1017/S0373463311000087 Special-Committee-159 (2001) Minimum operational performance
Groves PD (2013) Principles of GNSS, inertial, and multi-sensor inte- standards for global positioning system/wide area augmentation
grated navigation systems (GNSS technology and applications), system airborne equipment. Document, DO-229C, RTCA
2nd edn. Artech House Publishers, Boston Sun D, Petovello MG, Cannon ME (2013) Ultratight GPS/reduced-
Groves PD, Jiang Z (2013) Height aiding, C/N0 weighting and consist- IMU integration for land vehicle navigation. IEEE Trans Aerosp
ency checking for GNSS NLOS and multipath mitigation in urban Electron Syst 49:1781–1791
areas. J Navig 66:653–669 Suzuki T, Kubo N (2013) Correcting GNSS multipath errors using
Groves PD, Wang L, Adjrad M, Ellul C (2015) GNSS shadow match- a 3D surface model and particle filter. In: Proceedings of ION
ing: the challenges ahead. In: Proceedings of ION GNSS + 2015, GNSS + 2013, Nashville, Tennessee, USA, 16–20 Sept 2013,
Tampa, Florida, Sept 2015, pp 2421–2443 pp 1583–1595
Gu Y, Hsu L-T, Kamijo S (2015) GNSS/on-board inertial sensor Veitsel VA, Zhdanov AV, Zhodzishsky MI (1998) The mitigation of
integration with the aid of 3D building map for lane-level vehi- multipath errors by strobe correlators in GPS/GLONASS receiv-
cle self-localization in urban canyon. IEEE Trans Veh Technol ers. GPS Solut 2:38–45
65:4274–4287 Wang JH, Gao Y (2010) Land vehicle dynamics-aided inertial naviga-
Han H, Wang J, Wang J, Tan X (2015) Performance analysis on carrier tion. IEEE Trans Aerosp Electron Syst 46:1638–1653
phase-based tightly-coupled GPS/BDS/INS integration in GNSS Wang L, Groves P, Ziebart M (2012) Multi-constellation gnss perfor-
degraded and denied environments. Sensors 15:8685 mance evaluation for urban canyons using large virtual reality city
Hartinger H, Brunner FK (1999) Variances of GPS phase observations: models. J Navig 65:459–476
the SIGMA-ɛ model. GPS Solut 2(4):35–43 Wang L, Groves PD, Ziebart MK (2013) GNSS shadow matching:
HKPSG (2016) Hong Kong planning standards and guidelines. Tech- improving urban positioning accuracy using a 3D city model with
nical Services Section of Planning Department, Government of optimized visibility scoring scheme. Navigation 60:195–207
Hong Kong, Wan Chai Wang L, Groves PD, Ziebart MK (2015) Smartphone shadow matching
Hsu L-T, Jan S-S, Groves P, Kubo N (2015) Multipath mitigation and for better cross-street GNSS positioning in urban environments.
NLOS detection using vector tracking in urban environments. J Navig 68:411–433
GPS Solut 19:249–262 Yozevitch R, Ben-Moshe B (2015) A robust shadow matching algo-
Hsu L-T, Gu Y, Huang Y, Kamijo S (2016a) Urban pedestrian naviga- rithm for GNSS positioning. Navigation 62:95–109
tion using smartphone-based dead reckoning and 3D maps aided Yozevitch R, Ben-Moshe B, Dvir A (2014) GNSS accuracy improve-
GNSS. IEEE Sens J 16:1281–1293 ment using rapid shadow transitions. IEEE Trans Intell Transp
Hsu L-T, Gu Y, Kamijo S (2016b) 3D building model-based pedestrian Syst 15:1113–1122
positioning method using GPS/GLONASS/QZSS and its reliabil- Yozevitch R, Ben-Moshe B, Weissman A (2016) A robust GNSS LOS/
ity calculation. GPS Solut 20:413–428 NLOS signal classifier. Navigation 63:429–442
Hsu L-T, Tokura H, Kubo N, Gu Y, Kamijo S (2017) Multiple faulty Zhdanov A, Zhodzishsky M, Veitsel V, Ashjaee J (2002) Evolution
GNSS measurement exclusion based on consistency check in of multipath error reduction with signal processing. GPS Solut
urban canyons. IEEE Sens J 17(6):1909–1917 5:19–28
Isaacs JT, Irish AT, Quitin F, Madhow U, Hespanha JP Bayesian locali-
zation and mapping using GNSS SNR measurements. In: 2014
IEEE/ION position, location and navigation symposium—PLANS Li‑Ta Hsu  received the B.S.
2014, 5–8 May 2014, pp 445–451 and Ph.D. degrees in aeronautics
Iwase T, Suzuki N, Watanabe Y (2013) Estimation and exclusion of and astronautics from National
multipath range error for robust positioning. GPS Solut 17:53–62 Cheng Kung University, Taiwan,
Jiang Z, Groves P (2014) NLOS GPS signal detection using a dual- in 2007 and 2013, respectively.
polarisation antenna. GPS Solut 18:15–26 He is currently an assistant pro-
Kanwal N, Hurskainen H, Nurmi J (2010) Vector tracking loop design fessor with Interdisciplinary
for degraded signal environment. In: Ubiquitous positioning Division of Aeronautical and
indoor navigation and location based service, Kirkkonummi, Aviation Engineering, The Hong
Finland, 14–15 Oct 2010, pp 1–4 Kong Polytechnic University,
Kumar R, Petovello MG (2014) A novel GNSS positioning technique before he served as post-doctoral
for improved accuracy in urban canyon scenarios using 3D city researcher in Institute of Indus-
model. In: Proceedings of ION GNSS + 2014, Tampa, Florida, trial Science at University of
USA, 8–12 Sept 2014, pp 2139–2148 Tokyo, Japan. In 2012, he was a
Li B, Cui W, Wang B (2015) A robust wireless sensor network localiza- visiting scholar in University
tion algorithm in mixed LOS/NLOS scenario. Sensors 15:23536 College London, U.K. His
Misra P, Enge P (2011) Global positioning system: signals, measure- research interests include GNSS positioning in challenging environ-
ments, and performance. Ganga-Jamuna Press, Lincoln, p 01773 ments and localization for autonomous driving vehicle and unmanned
Miura S, Hsu LT, Chen F, Kamijo S (2015) GPS error correction with aerial vehicle.
pseudorange evaluation using three-dimensional maps intelligent.
IEEE Trans Transp syst 16:3104–3115

13