Lane Accurate Detection of Map Changes

based on Low Cost Smartphone Data

Florian Jomrich

1,2

, Daniel Bischoff

1,2

, Steffen Knapp

, Tobias Meuser

, Bj

orn Richerzhagen

and Ralf Steinmetz

Opel Automobile GmbH, 65423 R

usselsheim, Germany

Multimedia Communications Lab (KOM), TU Darmstadt, 64283 Darmstadt, Germany

Keywords:

Map Change Detection, Low Cost, Smartphones, Sensor Fusion, Lane Change Detection.

Abstract:

Self-driving vehicles rely on High Deﬁnition Street Maps (HD Maps) to ensure the safety and comfort of their

driving capabilities. However, the road network infrastructure is subject to constant changes (e.g. through

constructions works, accidents, ...). Such changes have to be quickly identiﬁed to avoid dangerous driving

situations, for example through a reduction of driving speed or the safe handover of driving control back to

the human. To address this issue we propose a road hazard detection algorithm that identiﬁes and marks the

extent of such changes based on crowdsourced GNSS data. To increase the detection speed of our proposed

algorithm, we only rely on sensor information in the collection process, that is not only available through

vehicles, but as well by cheap and ubiquitous devices carried on by the passengers such as smartphones. To

deal with the limited accuracy of the collected data, we enhance existing algorithmic clustering approaches

by leveraging additional meta-data such as the quality of the collected GNSS points and the vehicle’s current

lane position. Our concept is evaluated with real world measurements in a highway construction site scenario

showing improved performance in comparison to the Kernel Density Estimation reference algorithm, used

versatile in Related Work.

1 INTRODUCTION

Highly automated vehicles are currently a very promi-

nent research topic (Brenner and Herrmann, 2018).

To enable the respective functionality, cars rely on

a large variety of sensors like cameras, radar, ul-

tra sonic sensors or lidar (Ziegler and et al., 2014).

Given the complexity of certain trafﬁc situations and

the limitations of individual sensors, current systems

further rely on a High Deﬁnition Street Map (HD

Map)(Madrigal, 2014; Miller, 2014). The HD Map

can be seen as an additional virtual sensor. It im-

proves the performance and accuracy of the car’s lo-

calization and classiﬁcation capabilities (e.g., of road

signs). It further enables the car to anticipate upcom-

ing street curvatures, which are not yet detectable by

the on-board sensors. This enhances safety and com-

fort for the passengers, as the vehicle can compare

its own sensor readings with the digital reference pre-

sented by the map.

To initially create HD Maps, specialized vehi-

cles with expensive measurement equipment are used.

These cars can rely on costly, high-precision sensors

like Differential GPS (DGPS) and laser scanners to

obtain precise measurement data.

However, the road infrastructure is permanently

changing (Plack, 2013). Highly automated vehicles

are only able to perform correct driving manoeuvres

based on correct, up to date map material. This makes

old and outdated information in the HD map a very

challenging problem(Rabel, 2017).

Several approaches have already been proposed

to leverage the global navigation satellite system

(GNSS) data obtained from common production ve-

hicles to create or update standard navigation maps

(Br

untrup and et al., 2005; Nieh

ofer and et al., 2009;

Cao and Krumm, 2009; Davies and et al., 2006;

Ahmed and et al., 2015). However, these approaches

did not address the time critical conditions to pre-

cisely identify road hazards, necessary to ensure the

safety of self-driving cars requiring an update of their

HD map material.

To close this research gap, we argue that any ad-

ditional sensor device that is carried by the passen-

gers, capable of collecting GNSS data, should be con-

sidered in the detection process of outdated HD Map

126

Jomrich, F., Bischoff, D., Knapp, S., Meuser, T., Richerzhagen, B. and Steinmetz, R.

Lane Accurate Detection of Map Changes based on Low Cost Smartphone Data.

DOI: 10.5220/0007709401260137

In Proceedings of the 5th International Conference on Vehicle Technology and Intelligent Transport Systems (VEHITS 2019), pages 126-137

ISBN: 978-989-758-374-2

material. As such devices are ubiquitously carried by

humans, they can tremendously increase the detection

speed of road hazards to ensure the overall driving

safety of self-driving vehicles. Therefore, we rely in

our presented work only on low precision GNSS sen-

sor readings obtained by smartphones carried inside

of a vehicle, as an example of such ubiquitous de-

vices. We further use readings from additional low

cost sensors available in the smartphones as meta-

information to enhance the overall precision of the

collected GNSS data. Based on this additional meta-

information we present a novel weighting algorithm

for the clustering procedure of GNSS data. Our algo-

rithm outperforms existing approaches in the creation

of lane-accurate clustering results, as evaluated with

real-world measurements in a highway construction

site scenario.

Based on the proposed weighting algorithm, we

present our second contribution, an algorithm that is

able to quickly identify lane accurate deviations in the

road network infrastructure. The detection is based on

the deviation between former obtained clustering re-

sults and the latest collected data as well as the overall

deviation in the average speed of the vehicles between

those two states over time. The performance of the al-

gorithm is evaluated in the aforementioned scenario.

The identiﬁed areas in the outdated HD map thus can

be safely avoided by handing over the driving con-

trol back to the human driver. Furthermore the iden-

tiﬁed HD Map areas can be updated more quickly, by

speciﬁcally requesting precise sensor readings from

the measuring vehicles for them.

After this introduction our paper is structured as

follows. In Section 2 we discuss related work for map

generation and updating from crowdsourced GNSS

traces and the utilization of low cost sensors to im-

prove localization. Based on our survey of related

work, we present our concept for map change detec-

tion consisting of a clustering algorithm enhancement

relying on meta information (Sec. 3) and a deviation

detection algorithm (Sec. 4). The real world scenario

used to evaluate the performance of our proposed al-

gorithms is described in Section 5. We evaluate our

proposed algorithm in Section 6, verifying that the

consideration of meta-information achieves improved

clustering and detection performance compared to ex-

isting approaches. We ﬁnally conclude the paper in

Section 7.

2 RELATED WORK

In the following, we structure and discuss related

work for map creation and map updates according to

the achieved accuracy (road- or lane-level) and uti-

lized sensor equipment.

2.1 Road-level Map Generation with

Floating Car Data

(Br

untrup and et al., 2005) as well as (Nieh

ofer and

et al., 2009) propose a client/server-based architecture

that infers the road infrastructure given GNSS traces

of an unknown area. Niehofer et al. explicitly utilize

GNSS data collected with mobile phones, supporting

our own work. The traces are used to either (i) create

a completely new road in the network or (ii) gradually

update an existing road with further data points. (Cao

and Krumm, 2009) rely on a ﬂeet of taxis to acquire

their testing data to create road maps. Their approach

to infer route data from GNSS traces is innovative by

performing so called energy well calculations to as-

sign newly arriving trace data to existing clustering

results while altering that data as well. These con-

cepts of gradually updating the already obtained clus-

tering results, to save time, inspired the development

of our own deviation detection algorithm. (Davies and

et al., 2006) present a framework to create road accu-

rate maps out of collected ﬂoating car data. The au-

thors state that modern day GNSS receivers (based on

the publications of (K.D. McDonlad and C. Hegarty,

2000) and (R. Prasad and M.Ruggieri, 2005)) have an

average standard deviation σ between 3.5 and 4.5 me-

ters. They further state that based on the central limit

theorem more than 70 GNSS traces would need to be

collected to differentiate between two adjacent roads

(not lanes as required for self-driving vehicles). Our

work aims to avoid these limits by not only relying on

GNSS trace data alone, but on fusing the readings of

several low cost sensors together with it.

A comparison of the aforementioned approaches

and further works is presented by (Ahmed and et al.,

2015). The authors provide OpenStreetMap bench-

mark material and evaluate the algorithms’ runtime,

ranging from several minutes to several hours. Their

results motivate the necessity for the development of

performance optimized map updating algorithms.

All of the aforementioned approaches only

achieve road-level accuracy, relying solely on GNSS

data. To identify temporary changes in the road

infrastructure—e.g., construction sites— to ensure

the safety of self-driving vehicles, at least lane-level

accuracy is required.

2.2 Lane Accurate Map Generation

One of the ﬁrst concepts for the automated creation

of lane accurate street maps was proposed by (Betaille

Lane Accurate Detection of Map Changes based on Low Cost Smartphone Data

127

(a) Common approach

(b) Weighted approach

Figure 1: Illustration of lane center point calculation ap-

proaches. In our weighted approach points accuracy are in-

dicated by colour and radius.

and Toledo-Moreo, 2010). In their work they focus on

the deﬁnition of a lane accurate map and its general

creation procedure. To achieve the technically best

possible results the authors rely on the coupling of

high accurate kinematic GPS (PPK) with dead reck-

oning estimations of the vehicle itself. The required

precise hardware is resembling more the costly equip-

ment used in the aforementioned specialized measur-

ing vehicles to create the HD maps. It is out of the

scope of our personal work, as we focus to quickly

detect changes in the HD map material using widely

available sensor data.

(Chen and Krumm, 2010) propose an algorithm

to infer lane accurate road networks from a multitude

of GNSS traces through the use of Gaussian Mixture

Models. The authors preprocess their GNSS traces by

virtually segmenting them through equidistant (e.g.

each 50 meters) perpendicular lines across all lanes

in the driving direction of the vehicles (see Fig. 1a).

Therefore, a random trace is selected as reference and

segmented in equidistances. All other traces are then

cut accordingly. The resulting intersection points of

all traces are then used for the following clustering

procedure. This initial concept is also used simi-

larly in several other works(Uduwaragoda and et al.,

2013),(Neuhold and et al., 2017),(Sato and et al.,

2012). Consequently, we use it as a well established

foundation for our own approach described in Sec-

tion 3.2. Chen et al. cluster the intersection points us-

ing Gaussian Mixture Models to identify center points

of individual lanes. To test their approach the au-

thors collected data from a ﬂeet of 55 vehicles, each

equipped with a standard GPS logger comparable to

today’s chips built into smartphones. Although the

authors do not investigate changes of road networks

in their work, they state it as necessary future work,

motivating our own contribution.

More recent approaches closely related to our

own contribution are presented in (Uduwaragoda

and et al., 2013) and (Neuhold and et al., 2017).

Uduwaragoda et al. propose the usage of the Ker-

nel Density Estimation Algorithm (KDE) to identify

the center lanes of a common 4-5 lane width high-

way. They test their concept based on trace data that

has been collected from a ﬂeet of vehicles carrying

GPS enabled phones. The authors state that at least

150 traces collected by their vehicles are necessary to

achieve a suitable lane detection accuracy in their sce-

nario. (Neuhold and et al., 2017) achieve similar re-

sults using KDE in different scenarios and with differ-

ent GPS loggers. They still have to rely on a large set

of traces (80 − 200) for accurate lane detection. Fur-

thermore, they utilize information on legally required

distances between lanes to increase the performance

of their algorithm.

An approach that relies on the fusion of different

sensors is presented by Guo et al. (Guo and et al.,

2016). They use information from a low-cost GNSS

sensor, an inertial measurement system, and ortho-

graphic images provided by the on-board camera to

create lane speciﬁc graphs as foundation for a com-

plete map. Massow et al. (Massow and et al., 2016)

also signiﬁcantly enhance the performance of a pure

GNSS-based approach by using additional sensors

like a camera and a radar. Both, (Guo and et al., 2016)

and (Massow and et al., 2016) support the concept of

utilizing additional sensors to improve the accuracy of

clustering approaches for lane detection. Radar and

camera sensors, however, are rather expensive and

also only available in a fraction of currently deployed

vehicles. In the following we present works that rely

on low cost sensor equipment to achieve accurate lo-

calization.

2.3 Lane Accurate Localization with

Low Cost Sensors

Two of the most sophisticated works regarding the

lane accurate localization of a vehicle using only low

cost hardware have been performed independently

from each other by Aly et al. (Aly and et al., 2015)

and Wu et al. (Wu and et al., 2016). The authors

rely on cheap accelerometer and gyroscope sensors,

which are present in production vehicles and mobile

devices as used in our work. Their proposed solu-

tions do not require knowledge of the initial start-

ing position of the car. Instead, a Markov localiza-

tion model or a Gaussian probability distribution is

used to keep track of all possible initial lane positions.

Through the accelerometer and the gyroscope sensors

the drivers behavior can be identiﬁed (e.g. perform-

ing a certain pattern of lane changes as presented in

Figure 2). Thereby, the probability of presence for

VEHITS 2019 - 5th International Conference on Vehicle Technology and Intelligent Transport Systems

128

each lane increases or decrease over time with only

a single remaining lane in the end. To further im-

prove this concept, Aly et al. suggest to use additional

knowledge in the form of so called bootstrap and or-

ganic anchors. Boostrap anchors rely on trafﬁc rules,

e.g., a right turn is normally performed on the right-

most lane. Organic anchors are conditions on the road

surface, e.g. a pothole, that can also be detected by

these two sensors. The appearances of such anchors

can then be linked to certain lanes. Both approaches

achieve a lane identiﬁcation rate of between 80% and

86%. Comparable work regarding the identiﬁcation

of steering manoeuvres has been proposed by Chen

et al. (Chen and et al., 2015). They compare their

detection algorithm with a camera-based approach.

The comparison shows the robustness of the detection

based on low cost sensors, as they are independent

from weather effects like sun blinding, rain, fog, and

the day and night cycle. In a similar work, Ahmed

et al. (Ahmed and et al., 2017) propose to utilize

the On-board Diagnostic Interface (OBD) for addi-

tional sensor information. However, this approach re-

quires additional dedicated hardware in the vehicle.

As discussed, existing approaches for lane-level accu-

racy either require dedicated—and potentially expen-

sive or unavailable—sensors (Guo and et al., 2016;

Massow and et al., 2016) or a large amount of traces

used in the clustering process (Davies and et al., 2006;

Uduwaragoda and et al., 2013). Both issues prevent

a fast detection of temporary situations such as con-

struction sites or accidents. Other approaches rely on

assumptions that are not fulﬁlled under such condi-

tions, such as the correlation between the centers of

the different lanes as utilized in (Neuhold and et al.,

2017). The utilization of additional, low cost sen-

sor data from mobile phones as proposed in (Liu and

et al., 2017; Aly and et al., 2015; Wu and et al., 2016)

is a promising direction to achieve the desired accu-

racy and speed of a map change detection algorithm.

Based on their work using a similar lane change algo-

rithm, we detail our concepts for enhanced clustering

and deviation detection in the following sections.

3 ENHANCED CLUSTERING

WITH META INFORMATION

As motivated in the previous section, we combine

GNSS measurements with low cost sensor data (meta

information) to achieve precise clustering results on

lane-level. We propose to use the following additional

information: (i) the number of visible GNSS satellites

(Sec. 3.1), (ii) an estimation of the position error pro-

vided by the mobile device (Sec. 3.2), and (iii) the

accelerometer and the gyroscope of the smartphones

to detect lane changes and derive the current lane po-

sition based on (Aly and et al., 2015; Wu and et al.,

2016) (Sec. 3.3).

3.1 Number of Satellites and Accuracy

Parameters

To achieve a proper position estimation a GNSS de-

vice requires at least four satellites to be visible. If

more satellites are visible the devices are able to fur-

ther improve the position estimate with information

from the additional satellites. Current GNSS devices

provide an accuracy estimation with each position re-

turned via their APIs (van Diggelen, 2007). The cal-

culation of this accuracy estimation is at the discretion

of the device manufacturer and, therefore, can dif-

fer. The API provided by Android smartphones used

in our work speciﬁes the position accuracy as a one

sigma (68%) reliability estimation of the horizontal

accuracy.

Consequently, the estimated position lies

with a probability of 68% within a circle of the indi-

cated accuracy radius around the estimated position.

In the following, we argue that a higher estimated

precision and a higher number of satellites corre-

lates with a position estimate of higher quality, which

should be considered more important for the overall

lane clustering procedure. We leverage this informa-

tion to achieve better clustering results in less time as

explained in the next section.

3.2 Proposed Weighted Clustering

Approach

To reduce the amount of GNSS traces that are re-

quired to properly identify changes on the lane-level,

we introduce a weighting factor into the common

clustering approach (Chen and Krumm, 2010). We

automatically annotate the measured GNSS position

estimates with the number of satellites and achieved

position accuracy for further consideration. In con-

trast to the common clustering approach(Chen and

Krumm, 2010), as explained in Section 2.2 (see Fig-

ure 1a as well), we do not select a random trace as

reference line. Instead, we select the trace that has the

highest number of satellites and accuracy as reported

by the mobile device. We assume that this trace will

mimic the overall road curvature best and therefore

improves the initial segmentation and the creation of

segment lines. As not all measured GNSS points are

directly located at a segmentation line, we generate

https://developer.android.com/reference/android/

location/Location.html

Lane Accurate Detection of Map Changes based on Low Cost Smartphone Data

129

artiﬁcial GNSS intersection points. As described by

Equation 1 the meta information values of those arti-

ﬁcial intersection points are calculated with respect to

the euclidean distance (d) of the closest preceding and

succeeding measured points and their actual values.

To complete the clustering process we then apply the

weighted mean as shown in Equation 2 for each of the

segments. The weighting function ( f

) is supposed to

be adapted accordingly to the available GNSS data.

In our evaluation in Section 6.3 we investigate differ-

ent weighting functions to ﬁnd the one with the best

performance for our data set. This way the latitude

and longitude values of the related GNSS estimation

are weighted depending on their overall quality. The

concept is illustrated in Figure 1b, where the overall

quality of the position estimates is indicated by their

radius and colour. We assume and verify in our eval-

uation that even with a smaller number of collected

traces more precise results are achieved, as the points

with overall better accuracy are taken into higher con-

sideration during clustering.

meta value

intersection point

(acc,no o f sattelites) =

meta value

pre

∗ d

pre

+ meta value

succ

∗ d

succ

pre

+ d

succ

(1)

clustercenter(lat,lon) =

∑

(lat,lon) ∗ f

(accuracy,no o f sattelites)

∑

allweights

(2)

3.3 Annotation of Lane Numbers in the

Traces

Even with the additional quality parameters the iden-

tiﬁcation of distinct lanes from noisy GNSS data is

still a difﬁcult task. Common clustering approaches

overcome this problem through the sheer amount of

data points (Uduwaragoda and et al., 2013) that are

collected over time. However, this is a serious is-

sue for the detection of temporary deviations as con-

sidered in our work. To address this issue, we uti-

lize a lane change detection algorithm as proposed

in the related work (see Sec. 2.3) utilizing only low-

cost, low-energy accelerometer and gyroscope sen-

sors. Thereby, we can rely on the estimated lane num-

ber as additional meta information annotated to the

GNSS data.

For our evaluation we implemented a basic lane

change detection algorithm, similar to the ones pro-

posed by Aly et al.(Aly and et al., 2015) and Wu et

al. (Wu and et al., 2016), from which we derive the

vehicles current lane number. The algorithm relies

Figure 2: Behaviour of the accelerometer sensor, when per-

forming lane changes.

on the combined sensor readings of the accelerome-

ter and gyroscope of the smartphone and detects lane

changes based on a static-offset gate of the measured

min and max values per time interval. An example ac-

celerometer reading based on our personal data set is

shown in Figure 2. The picture has been created with-

out any further noise ﬁltering. Thus it shows very well

the capabilities of such sensors to properly detect sev-

eral consecutive lane changes and accordingly iden-

tify the vehicles current lane. Therefore we see the

Related Works of Aly and Wu as a prerequisite for

our personal work. In this paper we showcase the fea-

sibility of such a lane change detection approach in

our context of map change detection. Therefore the

comparison to our manually annotated ground truth

as investigated in Section 6.2 is its main purpose. The

discussed results show, that even our rather simple ap-

proach already achieved a good detection rate. Fur-

ther improvements regarding the lane position detec-

tion as suggested by Aly and Wu probably will im-

prove the achievable clustering results even more.

With the identiﬁcation of lane changes and the de-

rived annotation of the car’s lane position, we enable

our algorithm to directly separate our collected GNSS

data regarding a speciﬁc lane. As evaluated in Section

6.1 this provides our proposed concept with a signif-

icant advantage over related clustering concepts, like

the Kernel Density Estimation algorithm, which do

not rely on lane information. Even over a longer pe-

riod of collection time we consider this an important

advantage for our own approach as it cancels out most

of the inﬂuence of present Gaussian noise in the col-

lected GNSS data.

4 ROAD INFRASTRUCTURE

DEVIATION DETECTION

Based upon the aforementioned enhancements of the

clustering procedure, we developed an algorithm to

reliably detect lane accurate changes in the road net-

VEHITS 2019 - 5th International Conference on Vehicle Technology and Intelligent Transport Systems

130

work (e.g. introduced through construction sites or

accidents) as the second contribution of this work.

The algorithm identiﬁes and highlights those areas,

where a certain deviation between historic and newly

generated clustering results is detected. The com-

plete procedure is illustrated in Figure 3. First of

all the track to be considered for the clustering pro-

cedure has to be segmented as described in Section

3.2. For each segment and each lane a distinct lane-

segment center point has to be calculated. The en-

tirety of all those center points represents the curva-

ture of the speciﬁc lane. Initially, the newly calculated

center point is marked as unreliable until a certain

amount of traces X is available for the clustering pro-

cedure to reduce the inﬂuence of faulty sensors and

the overall Gaussian noise. The exact number of re-

quired traces should be derived individually from the

amount of incoming traces over time and the accuracy

requirements of the clustering result. We investigate

it for our own dataset in Section 6.4. The ﬂowchart

of Figure 3 now describes the procedure how further

traces are added to the dataset. For each new trace,

the trace itself and the last X traces are considered as

input for the clustering process. The inﬂuence of the

last X traces thereby is weighed based on their col-

lection time as described by Formula 3. New traces

are considered more valuable and degrade in impor-

tance over time. The importance of a trace thereby is

also dependent on its quality values (accuracy, no of

satellites) combined in a weighting factor w

meta

∗ e





(time

trace x

−time

newest trace

)

degradation f actor





(3)

If the newly obtained clustering result deviates by

a threshold value T from the currently assumed lane

center point, a deviation is assumed to be present. In

our example evaluation described in Section 6.7 we

considered a value of 2/3 ∗ lanewidth for T , as the

width of a lane accepted by the road authorities can

be signiﬁcantly lower. We consider this value there-

fore as a ﬁrst proposal that is to be further optimized

in future work. The old center point is then added

to a vector of historic center points. If this history

storage already contains other old center points the

new clustering result is further compared with each

of them by calculating their deviation T . If for one

of them the achieved value is smaller than T it is as-

sumed to be the new lane center point. This way our

algorithm is able to handle situations in which a con-

struction site has been ﬁnished and the old road course

Figure 3: Deviation detection algorithm.

is present again, without starting the initial collection

process once again. If no ﬁtting historic lane centers

are present, the newly calculated clustering result is

assumed to be the new lane center point for its related

segment on the highway.

Besides the longitudinal deviation along a track

our algorithm also has to provide the latitudinal ex-

tend of the detected road hazard. Only through an

early warning ahead of the road hazard a safe han-

dover from the self-driving vehicles back to a hu-

man driver can be ensured. Our proposed algorithm

therefore additionally considers the average achieved

speed for each lane segment along the track to identify

the actual extent of the detected construction side or

similar road hazard. As the speed is normally gradu-

ally reduced before a construction side (e.g. from 120

km/h down to 80 or 60 km/h) and then increased again

to its former value it provides a good indicator for be-

gin and end of the construction site. The achieved

results for our evaluation are presented in Figure 11.

Our algorithm only indicates the presence of a con-

structions side if both detection conditions are met.

Therefore the average speed at the considered track

segment has to be suitably reduced (80 km/h or be-

low) and a certain deviation in the road course has to

be present. Otherwise a false detection could be pos-

sible. A reduction of speed for example could also be

induced by a common trafﬁc jam in the morning rush

hour.

5 SCENARIO DESCRIPTION

To investigate the performance of our proposed de-

viation detection algorithm in a real-world scenario,

Lane Accurate Detection of Map Changes based on Low Cost Smartphone Data

131

we performed several test drives, to collect data from

construction sites. The dataset, which we used for this

publication, contains over 1.934.000 GNSS points.

It will be published on github

to provide bench-

mark material for future algorithms in this ﬁeld of re-

search. Each data point is annotated with our consid-

ered meta-information (the vehicles current lane, the

GNSS position’s accuracy value and the number of

available satellites). Furthermore we manually tagged

the start and the end of each experienced construction

site to later rely on as ground truth. The measurement

points have been collected from April 2016 to Jan-

uary 2017. Within this time over a dozen construc-

tions sites have been added into the dataset, located

in urban, suburban and rural areas. To illustrate the

capabilities of our approach and the developed algo-

rithm we show the results accordingly for a construc-

tion site on the German highway A67 between the

cities of R

usselsheim and Darmstadt as illustrated by

Figure 11. GNSS traces have been collected in both

directions, with 369 traces in the direction to Darm-

stadt and 292 in the direction to R

usslesheim. For

the following graphs we present the achieved results

of the driving direction from R

usselsheim to Darm-

stadt. We ensured to have a sufﬁciently large amount

of traces for each of the lanes of this two lane highway

to be able to evaluate the beneﬁts of our lane accurate

ﬁltering of the GNSS data.

To resemble an actual large scale deployment of

our approach several different smartphones (Nexus 4,

Nexus 5, Blackberry Classic and Samsung Galaxy

S7) with different quality levels of GNSS sensors

have been used to collect the location data. These

phones therefore were placed on the dashboards of the

probe vehicles to resemble a possible usage scenario

of the end-customers.

6 EVALUATION

We selected the Kernel Density Estimation Algorithm

(KDE) as reference algorithm, as it is the most fre-

quently used and best performing algorithm through-

out the Related Work presented in Section 2. To com-

pare the results of the KDE clustering algorithm with

our approach, we manually annotated a reference line

located on the center dashed line of the highway using

satellite images (see Figure 4). To be able to safely

perform its driving task an automated vehicle has to

stay in the boundaries of its current lane. Therefore

we chose the deviation of the calculated lane center

lines provided by the investigated algorithms and this

https://github.com/ﬂorianjomrich/

construction side traces fjom

Figure 4: Reference line created from satellite images of the

center dashed line.

reference line in the middle of the two lane highway,

as our performance metric. This initial performance

evaluation, described in the following paragraphs was

executed on a section of the highway A67, where no

construction side was present. That way we could rely

in our evaluation on the rules of federal regulation in

Germany. They require each lane of our investigated

two lane highway to be exactly 3.75 meters wide.

ideal clustering algorithm therefore would achieve a

calculated deviation from all its predicted lane center

points to the reference line of 1.85 meters. These op-

timums are indicated by dashed lines in the following

plots. Figure 5 shows the achieved results for both

lanes of the highway and all our investigated algo-

rithm combinations. It is explained in detail in the

following.

6.1 Effect of the Lane Annotation

In a ﬁrst step we evaluated the inﬂuence of the lane in-

dex parameter for each collected GNSS point on the

ﬁnal clustering results. This index number enables

the ﬁltering of GNSS traces initially regarding their

speciﬁc lanes before the data is handed over to the

clustering algorithm. Therefore the lane changes dur-

ing the test drives had been automatically annotated

through our described algorithm (see Section 3.3), as

well as manually by the push of an according button

for a ground truth comparison. The results presented

in the plots of Figure 5 are based on all 369 traces

that have been collected for the driving direction from

usselsheim to Darmstadt on the A67. To avoid the

malicious inﬂuence of GNSS points that might have

been collected exactly during a lane change our data

ﬁltering algorithm neglected those points. The three

points, which are measured directly previous to, dur-

ing the lane change and directly afterwards are re-

moved from the lane speciﬁc data set. To compare

the achievable clustering performance we executed a

common mean center point calculation algorithm for

each segment line using our lane speciﬁc ﬁltered data.

As comparison the Kernel Density Estimation Algo-

rithm was executed on the same data, but without the

https://www.forschungsinformationssystem.de/servlet/

is/275112/

VEHITS 2019 - 5th International Conference on Vehicle Technology and Intelligent Transport Systems

132

Figure 5: Comparison of different algorithm clustering re-

sults for both lanes and all available traces.

lane ﬁltering.

For both lanes it is clearly visible that prior ﬁlter-

ing based on lane annotations (as described in Section

3.3) has a signiﬁcant impact on the obtained cluster-

ing accuracy. The variance of the obtained KDE cen-

ter points is much higher, than the simple mean cal-

culation approach executed on the pre-ﬁltered data.

For some instances, the center point calculated by the

KDE algorithm lies outside of the actual lane bound-

aries. This is most likely due to bad reception ac-

curacy of GNSS receivers in current smartphones,

which can reach several meters. The simple mean

center point calculation, which could rely on our lane

speciﬁc pre ﬁltered data, achieved much better vari-

ance values and stayed within the lane boundaries for

all 225 individual segments along our selected track.

Unexpectedly, the results obtained for the right lane

had a visible static offset to the middle of the road

towards the center line. The average of all estimated

center points on the left lane in contrast was (as ex-

pected) close to the actual center of the lane.

We investigated this phenomenon by executing

several drive tests (37 in total) for which the drivers

had been told to stay on the right lane for the whole

time of the track. The obtained results therefore are

shown in the lowermost plot of Figure 5. It is clearly

visible that this plot does not have the aforementioned

static offset. As a conclusion we suppose that the

difference in the obtained results is probably due to

the driving behavior of the test drivers and the col-

lection procedure of sensor data readings within our

Android smartphones. As the investigated scenario is

a two lane highway the right lane was mostly occu-

pied by slowly driving trucks. The test drives have

been conducted with common sedans from our ﬂeet,

therefore most of the time these vehicles overtook the

slower driving trucks. In conclusion the vehicle ei-

ther stayed consistently on the left lane or drove on the

right lane with more interruptions due to lane changes

to overtake a truck. Within the Android smartphone

a Kalman ﬁlter is already smoothing the obtained

GNSS traces. In our opinion, the Kalman ﬁlter is

unable to capture two immediately consecutive lane

changes correctly, leading to the displacement of the

reported GNSS locations. This is an aspect that has

to be considered in future work, as one might want

to ﬁlter the incoming GNSS trace data accordingly.

The speed of the vehicles (trucks are slower than nor-

mal cars in average) and the previous driving behavior

(identiﬁable through other sensors like the accelerom-

eter and the gyroscope) are probably good indicators

therefore. Consequently, we present results based on

data gathered on the left lane of the highway for the

remainder of this evaluation.

6.2 Algorithmic vs. Manual Annotation

As a second step in our investigation we wanted to

compare the achievable accuracy results of our algo-

rithm induced lane annotation compared to the man-

ual annotated ground truth. This test was performed

to ensure that our proposed concept is also capable

to be deployed in future devices without the require-

ment of any mandatory input from the user. We im-

plemented the algorithm to automatically detect the

lane changes of the vehicle while driving based on

the status changes of the accelerometer and the gy-

roscope built into the smartphone, as described in

Section 3.3. The achieved results of the algorithmic

annotation in comparison to the manually annotation

of lane changes for both lanes are shown in Figure

5. It is clearly visible that both approaches perform

comparably well. There is no clearly better perform-

ing approach as for the left lane the results obtained

from the manually annotation seem to be slightly bet-

ter, whereas on the right lane the algorithmic anno-

tated results tend to be a bit better. The obtained re-

sults are very promising, as they showcase that an ac-

tual deployment of such a system is certainly possi-

ble. We assume the achievable results could be fur-

ther enhanced by optimisations in the lane change de-

tection. Our implementation for example, which was

purely based on static thresholds, performed consid-

erably good and might be further enhanced with dy-

namic speed dependent changes of the used thresh-

olds for the lane change indicators.

6.3 Impact of Weighting Functions

As the next step to improve the performance of our

lane clustering process even further we introduced

the aforementioned weighting concept for each col-

Lane Accurate Detection of Map Changes based on Low Cost Smartphone Data

133

Figure 6: Comparison of different weighting functions (left

lane data).

lected GNSS point. The weighting algorithm takes

the included meta data into account to calculate the

importance of each data point before the actual clus-

tering procedure is executed. It considers the number

of satellites, that were visible to the GNSS receiver

of the smartphone, as well as the accuracy value that

is provided by the Android API for each measured

GNSS point. As more available satellites tend to pro-

vide a better localisation we took their number di-

rectly into account for the weighting. The accuracy

value was considered inversely important as a smaller

accuracy means a more precise localisation.

To investigate the importance of both meta data

values, we evaluated different weighting functions

(see Formula 2), with the results being shown in Fig-

ure 6. The obtained results showcase that both, the

number of satellites and the accuracy, improve the

achievable precision, when considered with high im-

portance. For the comparison we again took all GNSS

points into consideration for the driving direction be-

tween R

usselsheim and Darmstadt. We achieved the

best clustering performance with the parameters set

to weigh the Numbero f Satellites

or 2

Accuracy

. As

the conﬁguration of 2

Accuracy

achieved less variance

we considered it as our reference weighting function

for the following evaluation steps. However, we see

these results still as a ﬁrst investigative look and con-

sider further optimization potential in the weighting

procedure itself for future work.

6.4 Clustering Performance for

Different Numbers of Traces

As the required time to achieve a reasonable clus-

tering result is crucial for our use cases, we also in-

vestigated the behaviour of our proposed weighted

mean clustering algorithm when considering differ-

ent amounts of traces as input. The obtained results

Figure 7: Inﬂuence of different amounts of available input

traces on clustering performance (left lane data).

for the left lane of the highway A67 are presented in

Figure 7. From the different boxplots it is visible that

our weighted mean clustering approach reaches an ac-

curacy within the suggested lane boundaries reason-

able fast between 35 to 45 randomly selected traces.

By increasing the amount of random traces up to 90

we achieve only slight performance improvements in

comparison. Although the clustering approach clearly

beneﬁts from a much higher number of traces as

shown by Figure 5, we can state that a difference de-

tection with an accuracy in the regions of the lane

width can be achieved much quicker. The overall

achievable performance can also be improved by se-

lecting only high quality traces as described in Sec-

tion 6.6. Future devices with a higher measuring ac-

curacy should obviously further reduce this required

number of traces.

6.5 Inﬂuence of Weighting on Lane

Filtered Data

Another question that we addressed in our work is the

impact of our additional meta information. We specif-

ically wanted to clarify if the information required for

the weighting procedure (number of satellites and the

accuracy) is worthy to consider for an initially lane-

ﬁltered data set or if the lane ﬁltering itself is the only

important factor required to achieve better clustering

results. Therefore we let our three algorithms (mean,

weighed mean and KDE) run on our full lane-ﬁltered

data set. The achieved results are presented in Figure

As expected the results of the KDE clustering al-

gorithm also beneﬁted from the lane ﬁltering pro-

cess. However its achieved clustering performance is

comparable to the standard mean calculation, which

achieves a much better execution time performance

compared to the KDE. The weighed mean achieved

VEHITS 2019 - 5th International Conference on Vehicle Technology and Intelligent Transport Systems

134

Figure 8: Comparison between weighting and non-

weighting algorithms using all available pre-lane ﬁltered

traces (left lane data).

the best clustering results based on our collected

GNSS traces. We are well aware that other works

might have achieved better overall performance re-

sults as they are highly dependent on the used hard-

ware. However, based on our results we could ver-

ify the performance improvements achievable with

the consideration of meta-data, even when relying on

a largely varying set of measurement devices as de-

scribed in Section 5. Furthermore, we did not take any

dependencies between the two calculated lane center

points into account, as done by other Related Work

(Neuhold and et al., 2017). We did not rely on them

to achieve our results, as these conditions might not

hold true in our considered scenarios of construction

works and accidents.

6.6 Inﬂuence of Trace Selection

Motivated from our previous results, we investigated

the inﬂuence of the overall quality of the collected

traces. Therefore we conducted a test where we com-

pared the 70 most accurate (best) traces of our full

data set with another set of 70 randomly chosen traces

(see Figure 9). The amount of 70 best traces was se-

lected, as our investigation indicated that at around

this amount of traces the achievable accuracy satu-

rated in our clustering approach. Thus, we assumed

that a lot more traces would be required to further im-

prove the clustering results. As time is a critical factor

in the updating procedure of a map regarding a con-

struction side or an accident this aspect was critical

for us to be investigated. The best traces were selected

based on the average accuracy of all their interception

points with the created segment lines. The trace with

the lowest average in those 225 points is the best trace

of our data set. Figure 9 clearly shows that the selec-

tion of traces has a signiﬁcant impact on the quality of

Figure 9: Performance comparison between randomly se-

lected traces and the most accurate available traces (left lane

data).

the clustering result. As expected the selection of best

traces achieved a signiﬁcantly better clustering result

than the other group based on a random selection. In-

terestingly, the 70 selected traces achieved compara-

bly good performance results as using all traces that

we could collect (indicated by two upmosts plot in

Figure 9). This strengthens our initial assumption. As

a conclusion it might be well worthy to consider only

a high qualitative subset of all collected traces in a

continuous updating procedure over a long period of

time, with many cars driving by, to successfully and

efﬁciently maintain an accurate update of the current

status of the road network. If new traces do not show

any signiﬁcant sign of deviation from the current clus-

tering results or they are not much better in the av-

erage of their overall achieved accuracy values, they

might possibly be neglected as well. However, further

investigation in future work with a larger set of traces

is required to verify this observation.

6.7 Construction Site and Deviation

Detection

Based on the positive results obtained from the previ-

ously described evaluation steps we than further con-

tinued the evaluation of our deviation detection algo-

rithm. We evaluated the ﬁrst performance results of

our algorithm in an example scenario on a different

part of the highway A67, where a construction site

was present in May 2016, as illustrated by the Google

Earth satellite images in Figure 10 and 11. The inves-

tigated scenario shows the feasibility of our deviation

detection algorithm to react on updates of the road

structure very quickly as the provided data set for this

section of the highway with its speciﬁc construction

side status contained only a rather small data set of

15-25 GNSS traces per lane. Even with this small

Lane Accurate Detection of Map Changes based on Low Cost Smartphone Data

135

Figure 11: Correlation between the average vehicles speed (indicated by colour) and the location of the construction side

(indicated by triangles). Speeds right before and in the construction side range from 90 km/h down to 57 km/h (yellow - red).

Before and after the construction side normal driving speeds of around 125 km/h are reached in average (green).

(a) Start

(b) Middle

Figure 10: Sections of the investigated construction site

(yellow) in correlation with the situation after the comple-

tion (cyan).

amount of data our algorithm was able to resemble the

entrance and the course of the construction side well

and stays in the given lane boundaries in each consid-

ered segment. Only at the exit of the construction side

is an unrealistic assumed distance between the two

lane center points visible. We propose to avoid such

situations by relying only on a larger subset of traces,

as investigated for our dataset in Section 6.4. The in-

troduction of a minimum distance condition between

the two adjacent lanes in the post processing of the

GNSS data is considered as part of our future work.

A further reduction of the distance between two adja-

cent clustering segments in a detected deviation envi-

ronment could probably also improve the achievable

performance.

7 CONCLUSION AND FUTURE

WORK

Within this work we presented our deviation detection

algorithm to detect and visualize the extend of con-

struction sites and other road hazards on a lane accu-

rate level. This algorithm has been designed with the

usage scenario of frequent HD-Map updates in mind.

Such lane accurate maps are required to improve the

safety and comfort while driving highly automated.

To ensure the accuracy of such maps time critical map

updates are required.

Our deviation detection algorithm mainly bene-

ﬁts from improvements in the general clustering pro-

cess required for identifying the current lane curva-

ture. These are namely the usage of available meta-

information such as the number of available satel-

lites, the accuracy of the obtained GNSS location

and the current lane in which the vehicle was driv-

ing at that point in time. This meta-information is

used to improve the commonly known clustering pro-

cess of the collected GNSS traces as described in the

Related Work (Chen and Krumm, 2010). We eval-

uated the beneﬁts of this additional meta-data in a

comparison of our own processing pipeline with the

state of the art Kernel Density Estimation clustering

algorithm, showing signiﬁcant performance improve-

ments. The evaluation procedure was based on a large

self-collected GNSS data set, which was obtained

from common smartphones in a highway driving sce-

nario. We discuss and evaluate in this work that such

cheap mobile devices can provide lane-accurate lo-

cation information required for the HD-map update

through our proposed intelligent sensor fusion.

The proposed weighting algorithm of the initially

obtained GNSS data performs well but can be fur-

ther enhanced in future work. Possible improvements

affect the importance consideration of the incoming

trafﬁc data and the weighting of the quality of the

traces. This includes the consideration of inﬂuence

factors such as the quality of the used GNSS chipsets,

VEHITS 2019 - 5th International Conference on Vehicle Technology and Intelligent Transport Systems

136

which is likely to increase in the following years

, as

well as the adaptation of the clustering process reso-

lution based on the average driving speed of the vehi-

cles.

REFERENCES

Ahmed, M. and et al. (2015). A comparison and evaluation

of map construction algorithms using vehicle tracking

data. GeoInformatica, 19(3):601–632.

Ahmed, U. and et al. (2017). Minimizing gps depen-

dency for a vehicle’s trajectory identiﬁcation by us-

ing data from smartphone inertial sensors and onboard

diagnostics device. Transportation Research Record:

Journal of the Transportation Research Board, pages

55–63.

Aly, H. and et al. (2015). Lanequest: An accurate and

energy-efﬁcient lane detection system. In Perva-

sive Computing and Communications (PerCom), 2015

IEEE International Conference on, pages 163–171.

IEEE.

Betaille, D. and Toledo-Moreo, R. (2010). Creating en-

hanced maps for lane-level vehicle navigation. IEEE

Transactions on Intelligent Transportation Systems,

11(4):786–798.

Brenner, W. and Herrmann, A. (2018). An Overview

of Technology, Beneﬁts and Impact of Automated

and Autonomous Driving on the Automotive Indus-

try. In Digital Marketplaces Unleashed, pages 427–

442. Springer, Berlin, Heidelberg. DOI: 10.1007/978-

3-662-49275-8 39.

untrup, R. and et al. (2005). Incremental map generation

with GPS traces. In Intelligent Transportation Sys-

tems, 2005. Proceedings. 2005 IEEE, pages 574–579.

IEEE.

Cao, L. and Krumm, J. (2009). From GPS traces to a

routable road map. In Proceedings of the 17th ACM

SIGSPATIAL international conference on advances in

geographic information systems, pages 3–12. ACM.

Chen, D. and et al. (2015). Invisible sensing of vehicle

steering with smartphones. In Proceedings of the 13th

Annual International Conference on Mobile Systems,

Applications, and Services, pages 1–13. ACM Press.

Chen, Y. and Krumm, J. (2010). Probabilistic modeling of

trafﬁc lanes from GPS traces. In Proceedings of the

18th SIGSPATIAL International Conference on Ad-

vances in Geographic Information Systems, pages 81–

88. ACM.

Davies, J. J. and et al. (2006). Scalable, distributed, real-

time map generation. IEEE Pervasive Computing,

5(4):47–54.

Guo, C. and et al. (2016). A low-cost solution for automatic

lane-level map generation using conventional in-car

sensors. IEEE Transactions on Intelligent Transporta-

tion Systems, 17(8):2355–2366.

https://www.pocketnavigation.de/2017/10/

broadcom-gps-chip-zentimeter/

K.D. McDonlad and C. Hegarty (2000). Postmoderniza-

tion GPS Performance Capabilites. Proc. IAIN World

Congress and the ION 56th Ann. Meeting. Inst. of

Naviga- tion.

Liu, Z. and et al. (2017). A participatory urban trafﬁc mon-

itoring system: The power of bus riders. IEEE Trans-

actions on Intelligent Transportation Systems, pages

1–14.

Madrigal, A. C. (2014). The Trick That Makes Google’s

Self-Driving Cars Work - The Atlantic [online].

Massow, K. and et al. (2016). Deriving HD maps for highly

automated driving from vehicular probe data. In In-

telligent Transportation Systems (ITSC), 2016 IEEE

19th International Conference on, pages 1745–1752.

IEEE.

Miller, G. (2014). Autonomous cars will require a totally

new kind of map | WIRED [online].

Neuhold, R. and et al. (2017). Generating a lane-speciﬁc

transportation network based on ﬂoating-car data. In

Stanton, N. A., Landry, S., Di Bucchianico, G., and

Vallicelli, A., editors, Advances in Human Aspects

of Transportation, volume 484, pages 1025–1037.

Springer International Publishing.

Nieh

ofer, B. and et al. (2009). GPS community map

generation for enhanced routing methods based on

trace-collection by mobile phones. In Proeedings of

the First International Conference on Advances in

Satellite and Space Communications, pages 156–161.

IEEE.

Plack, J. (2013). The unmatched quality of HERE maps

content - HERE 360 [online].

R. Prasad and M.Ruggieri (2005). Applied Satellite Navi-

gation Using GPS, G ALILEO , and Aug- mentation

Systems. Artech House.

Rabel, D. (2017). How HERE HD live map perpetually

heals itself [online].

Sato, N. and et al. (2012). Estimating the number of lanes

on rapid road map survey system using GPS trajecto-

ries as collective intelligence. In Proceedings of the

15th International Conference on Network-Based In-

formation Systems, pages 82–88. IEEE.

Uduwaragoda, E. and et al. (2013). Generating lane level

road data from vehicle trajectories using kernel den-

sity estimation. In 16th International IEEE Con-

ference on Intelligent Transportation Systems (ITSC

2013), pages 384–391. IEEE.

van Diggelen, F. (2007). Update: GNSS accuracy: Lies,

damn lies, and statistics : GPS world.

Wu, Z. and et al. (2016). L3: Sensing driving conditions for

vehicle lane-level localization on highways. In IEEE

INFOCOM 2016-The 35th Annual IEEE International

Conference on Computer Communications, pages 1–

9. IEEE.

Ziegler, J. and et al. (2014). Making bertha drive - an au-

tonomous journey on a historic route. IEEE Intelligent

Transportation Systems Magazine, 6(2):8–20.

Lane Accurate Detection of Map Changes based on Low Cost Smartphone Data

137