Monocular Rear Approach Indicator for Motorcycles

Joerg Deigmoeller

, Herbert Janssen

, Oliver Fuchs

and Julian Eggert

Honda Research Institute Europe, Carl-Legien-Strasse 30, 63073 Offenbach, Germany

Honda R&D Europe (Germany), Carl-Legien-Strasse 30, 63073 Offenbach, Germany

Keywords:

Driver Assistance System, Monocular Camera, Optical Flow.

Abstract:

Conventional rear-view mirrors on motorcycles only allow a limited visibility as they are shaky and cover a

small ﬁeld of view. Especially at high speeds with strong headwind, it is difﬁcult for the rider to turn his head

to observe blind spots. To support the rider in observing the rear and blind-spots, a monocular system that

indicates approaching vehicles is proposed in this paper. The vision based indication relies on sparse optical

ﬂow estimation. In a ﬁrst step, a rough separation of background and approaching object pixel motion is done

in an efﬁcient and computationally cheap way. In a post-processing step, pixel motion information is further

checked on geometric meaningful transformations and continuity over time. As a prototype, the system has

been mounted on a Honda Pan-European motorcycle plus monitor in the dashboard that shows the rear-view

image to the rider. If an approaching object is detected, the rider gets an indication on the monitor. The rear-

view on the monitor not only acts as HMI (Human Machine Interface) for the indication, but also signiﬁcantly

extends the visibility compared to mirrors. The algorithm has been extensively evaluated for relative speeds

from 20 km/h to 100 km/h (speed differences between motorcycle and approaching vehicle), at normal, rainy

and night conditions. Results show that the approach offers a sensing range from 20 m at low speed up to 60

m at night.

1 INTRODUCTION

On motorcycles, the visibility to all sides is signiﬁ-

cantly worse compared to cars. The rear-view is only

possible by means of two mirrors, instead of three,

whereas the ﬁeld of vision is partially hidden by the

riders body. Additionally, rear-view mirrors have a

tendency to tremble at high velocity or on uneven

ground. Therefore, improved rear-view and rider as-

sistance are of great importance in the motorcycle do-

main to reduce the number of accidents and fatalities.

So far, there exist only special rear view sys-

tems (Quick, 2012) and blind spot assistance systems

for cars. Latter mainly make use of radar ((Audi,

2013),(Mazda, 2011),(Hella, 2012)) or sonar (Bosch,

2012) sensors. Both sensors are not well suited for

motorcycles as they require further treatment or even

additional sensors like gyroscopes to deal with lean-

ing conditions. Additionally, radar is quite heavy and

expensive compared to the overall costs of a motorcy-

cle. Sonar, in turn, only has a range of up to 6 m (at

least for an acceptable sensor size). Therefore, cam-

eras represent a good alternative with respect to costs,

size and sensing range. Despite that, cameras can

cope with leaning conditions and provide a rear-view

image that can be displayed on a dashboard monitor.

Existing work on camera-based assistance sys-

tems are unfortunately only available for cars (Nis-

san, 2012). Commonly, the scaling factor of detected

objects in consecutive video images, i.e. its change of

size overtime is used to decide whether it is approach-

ing or not. For the detection of objects, (Stierlin and

Dietmayer, 2012) and (Mueller et al., 2008) use op-

tical ﬂow information, whereas (Stein et al., 2003))

apply an appearance based method.

More complex approaches using pixel motion in-

formation from monocular images requires an ego-

motion-compensation at ﬁrst to detect other moving

road users (Ma et al., 2004). This can either be

done by feeding data from an IMU (Inertial Measure-

ment Unit) and reﬁning the ego-motion based on vi-

sual motion information (Rabe et al., 2007) or relying

on vision information only ((Klappstein et al., 2006),

(Scaramuzza and Siegwart, 2008)). Those methods

are mainly used for front-facing cameras.

As soon as the camera is mounted to the rear,

motion segmentation becomes much easier, as back-

ground motion caused by the ego-vehicle and object

motion caused by approaching objects signiﬁcantly

differ in their scaling factor (see discussion in follow-

474

Deigmoeller J., Janssen H., Fuchs O. and Eggert J..

Monocular Rear Approach Indicator for Motorcycles.

DOI: 10.5220/0004668304740480

In Proceedings of the 9th International Conference on Computer Vision Theory and Applications (VISAPP-2014), pages 474-480

ISBN: 978-989-758-009-3

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

ing chapter), because the former is contracting while

the latter is expanding.

Therefore, the method developed within this work

is also based on monocular pixel motion. The deci-

sion was against a stereo system to save one camera.

Additionally, the baseline of a stereo system would be

quite small because of the limited space on a motor-

cycle. This in turn restricts the sensing range signif-

icantly. Another advantage of using motion features

is the independence of an objects appearance, e.g. at

night the appearance changes signiﬁcantly compared

to day-time as a vehicle can only be identiﬁed by its

front-lights.

To the knowledge of the authors, there exists no

such vision-based system for motorcycles yet. As the

main difference in vehicle dynamics between car and

motorcycle is the ability to ride in leaning position,

(Schlipsing et al., 2012) proposed a method to esti-

mate the roll angle of a motorcycle. The idea is to

transfer existing assistance systems from the car do-

main, like lane detection or obstacle detections which

requires such a roll-angle compensation. Obviously,

this represents a convenient solution to make use of

already available technologies. The disadvantage is

that all post-processing depends on the reliability of

the roll-angle compensation, which might not be de-

sirable in sense of error propagation and independent

running applications.

In the remainder of this paper, the approaching ve-

hicle indication is described at ﬁrst in Section 2. The

implementation of the system on a motorcycle and ex-

perimental results under rainy, dark and high speed

conditions are discussed in Section 3. Finally, the dis-

cussion and conclusion section summarizes the out-

comes and explains remaining challenges.

2 APPROACHING VEHICLE

DETECTION

Mounting a camera to the rear of a vehicle causes a

contracting pixel motion in the image sequence if the

vehicle moves forward. This means that all pixels

move towards a focus of contraction if the scene is

static. The magnitude of each motion vector in the

image mainly depends on the corresponding depth of

a pixel in real world coordinates. If an object is mov-

ing in the scene, the measured pixel motion of the ob-

ject is a combination of the ego-motion and the object

motion. This results in zero motion if the object is

moving with the same speed in the same direction as

the ego-vehicle.

As soon as the object velocity is larger than the

ego-vehicle velocity, the pixel motion pattern of the

object becomes upscaling with a ﬂow ﬁeld mov-

ing away from a focus of expansion. This means,

during ego-vehicle movement, an approaching ob-

ject causes an expanding ﬂow ﬁeld while static back-

ground causes a contracting ﬂow ﬁeld. This makes

both patterns distinguishable by their scaling factor

(greater or lower than 1). In turn, if the object drives

with lower speed as the ego-vehicle, background and

object motion are both contracting, i.e. both scaling

factors are lower than 1.

If the ego-vehicle additionally undergoes a rota-

tional motion, the projected motion pattern on the im-

age plane is overlayed with a vertically translating

component in case of pitching or a rotational com-

ponent in case of rolling, whereas the magnitude of

a motion vector is independent of the corresponding

depth of a pixel. However, scaling factors are not in-

ﬂuenced by rotational motion. The rolling component

is of special interest for this application, as such a mo-

tion occurs only for motorcycles and not for cars.

In the following, potentially approaching object

motion is detected by simply checking the scale fac-

tor in a local neighborhood. This is an efﬁcient pre-

processing step to reduce the number of non-relevant

motion vectors for ﬁtting a geometric motion model

for approaching vehicles in a second step. The main

advantage of this approach is that no ego-motioncom-

pensation needs to be done at all, which avoids the

inﬂuence of errors from an additional pre-processing

step.

2.1 Pre-selection of Motion Information

For motion estimation, a sparse pixel motion estima-

tion method has been applied. Sparse means that mo-

tion vectors

= (u

)

with corresponding homo-

geneous pixel coordinates x

= (x

,1)

are only

computed at well structured regions. Compared to

dense motion estimations, which compute motion

vectors for every pixel, such methods can save signif-

icant computational effort. The method used here is

the pyramid implementation of the Lucas and Kanade

optical ﬂow estimation (Bouguet, 2000), available in

the OpenCV library (OpenCV, 2013). The big advan-

tage of the pyramid implementation compared to the

standard Lucas and Kanade approach is the ability to

cover large pixel displacements by propagating over

different image resolutions.

To decide whether a motion vector corresponds to

an approaching vehicle or to the background, at least

two neighboring motion vectors are required to com-

pute their scaling factor. As opposed to dense motion

vector ﬁelds, the neighborhood within a sparse mo-

tion vector ﬁeld is not clearly deﬁned.

MonocularRearApproachIndicatorforMotorcycles

475

motion

estimation

Delaunay

triangulation

pre-selection

Figure 1: Pre-selection of motion vectors that correspond

to an approaching object, starting with motion estimation,

followed by triangulating the coordinates of motion vectors

and, ﬁnally, keeping only those vector triples (green edges)

where all possible combinations of vectors fulﬁll s

> 1 and

> 1.

Therefore, a Delaunay triangulation (Shewchuk,

2002) is applied to create neighborly relations be-

tween all x

in a mesh. For triangulation, the software

Triangle by Jonathan Shewchuk is used (Shewchuk,

1996).

The Delaunay triangulation has the speciﬁc prop-

erty that within a circle that is drawn around three co-

ordinates, a triangle does not contain any other co-

ordinate of the complete mesh (see middle image in

Fig. 1). Such a network allows to compare each mo-

tion vector with its closest neighborswithin a triangle.

To make a decision whether a triple of motion vectors

may correspond to an approaching vehicle or not, the

scaling factor of two motion vectors at each edge of a

triangle is computed.

It is assumed that the motion in a close neighbor-

hood mainly consists of a translation T and scaling S,

whereas rotational and shearing as well as perspective

transformations can be neglected. A motion vector

at the homogeneous coordinate x

can be expressed as

follows:

= (A

− E

′

)

{z }

′

, where (1)



S T



, S =



0 s



T =





and E

′



1 0 0

0 1 0



′

has been subtracted here to allow a direct map-

ping between homogeneous coordinates x

and 2D-

vectors

. To estimate the scaling factors s

and s

two motion vectors within a triple are subtracted to

get rid off the translation T:

−

= A

′

− A

′

= A

′





− x

− y





(2)

Rearranging the equation above and solving for s

and s

yields:





−u

−x

+ 1

−v

−y

+ 1

(3)

Three motion vectors within a triangle are only

considered for further processing if all possible com-

binations of vector pairs fulﬁll the constraint to be up-

scale, i.e. s

> 1, s

> 1. Fig. 1 illustrates all steps for

pre-selecting motion vectors.

2.2 Geometric Model Fitting

After the pre-selection step, it is assumed that if any

motion information remains, it mainly corresponds

to approaching vehicles. Due to wrong or imprecise

measurement it is still possible that some motion vec-

tors may fulﬁll the constraint to be upscaling even if

they do not follow a meaningful motion.

Therefore, the widely used robust regression

method RANSAC (RANdom SAmple Consensus,

(Fischler and Bolles, 1981), (Ma et al., 2004)) is ap-

plied to ﬁt a geometric model into the remaining mo-

tion vectors to further check for a meaningful trans-

formation. The chosen model is an afﬁne transfor-

mation A, which is a good approximation when real

world coordinates lie on a plane parallel to the image

sensor and move towards the camera:

′

= (A− E

′

(4)

Again, E

′

has been subtracted to allow an afﬁne

mapping between x

and

′

. The RANSAC method

VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications

476

ﬁnds a model that supports as many motion vectors as

possible, which is deﬁned by the number of motion

vectors in the inlier set or also called consensus set

C(A):

C(A) = {

ε V : min

′

εM(A)

dist(

′

) ≤ 1px}, (5)

where V is the whole data set of motion vectors,

M(A) is the manifold of the model A and dist(·,·) is the

Euclidean distance between measured vector

and

model vector

′

. Here, the Euclidean distance is ﬁxed

to an accuracy of 1 pixel.

After ﬁtting the parameters of the afﬁne model,

A is decomposed into its components to identify

whether the motion pattern is upscaling or not:

A =



K T



, where (6)

K =





, and T =





The translational component T can be read off di-

rectly. In turn, scaling components can be separated

from rotational components, if shearing components

are comparably small within K, by taking the square

root of the diagonal entries of K

K ≈ (RS)

RS = S

= S

ES =



0 s



(7)

where S, R and E are 2 by 2 scaling, rotation and

identity matrices, with R

R = E. Only afﬁne motion

patterns that are upscaling are considered for post-

processing.

In this application, there remain two possible mo-

tion patterns after the pre-selection step. First, ap-

proaching vehicles, i.e. motion patterns that are of in-

terest. Second, if the motorcycle is standing still, e.g.

at a trafﬁc light. This type remains because the scal-

ing factor of the background motion hovers around

one (s

≈ 1, s

≈ 1) due to noise. Therefore, this

motion information can not be identiﬁed in the pre-

selection step based on local relations only. Instead,

the global afﬁne transformation clariﬁes this circum-

stance by its averaged scaling parameters. It has to be

mentioned, that in case of a non-moving motorcycle,

the static scene becomes planar and can be expressed

by an afﬁne transformation. Only if the scaling fac-

tors of the global transformation are below a certain

threshold t

, it is assumed that the motorcycleis stand-

ing still.

The RANSAC method is iteratively applied two

times to cover both possible motion patterns of an

approaching vehicle and the non-moving motorcycle.

If the ﬁrst ﬁt describes the situation of a motorcycle

which is standing still or s

and s

is even downscal-

ing, corresponding motion vectors are removed and a

second model is computed. In turn, if the ﬁrst ﬁt con-

tains an upscaling model, no more iteration is done.

The following pseudo-code illustrates this procedure:

FOR i = 0 to 1

doRANSAC() // compute afﬁne model

IF ((s

> t

) && (s

> t

))

removeOutlier() // remove motion vectors

// that do not ﬁt to model

break // break loop and do

// post-processing

ELSE

removeInlier() // remove motion vectors

// that ﬁt to model

continue // do second ﬁtting

ENDFOR

2.3 Post-processing

To make the detection of approaching vehicles more

robust, a temporal ﬁltering is applied. To do so, all

values of a 2-dimensional grid I

,t), with size

of the input image, are initialized with zero values.

Values at coordinates (x

)

are set to 1, if they re-

late to a motion vector in the current consensus set at

time t (see Fig. 2):

,t) = {

1 if

′

ε M(A)

0 else

(8)

,t) is then combined with values I(x

′

,t)

which have been predicted from previous time steps:

I(x

,t) = α I

,t) + (1 − α) I(x

′

,t), (9)

where α describes the decay over time.

In this application, α is set to 0.1 to make the ﬁl-

tering robust against outliers. To also allow a variance

in position, each coordinate value is blurred by a 3 by

3 Gaussian kernel followed by a median ﬁlter of size

3 by 3 to remove noisy areas.

An indication is ﬁnally given to the rider if the sum

of values in the ﬁltered 2-dimensional grid I(x

,t)

exceeds a certain threshold t

, i.e.

∑

med[G∗ I(x

,t)], (10)

where G is the Gaussian kernel and med is the me-

dian ﬁltering.

The prediction of coordinate values I(x

,t) to

the next time step I(x

′

,t + 1) is done with the aid

of the afﬁne transformation A from Equation 6:

MonocularRearApproachIndicatorforMotorcycles

477

predicted and accumulated coordinate values

coordinate values of current time step

Figure 2: Illustration of coordinate values that correspond

to the detected motion vectors (top and middle image) and

prediction plus accumulation over time (bottom image).



′



= K





+ T (11)

3 PROTOTYPE EVALUATION

The prototype motorcycle is a Honda Pan European

with a PlayStation Eye camera (75

◦

diagonal ﬁeld of

view, 640x480 at 15 frames per second) mounted to

the rear. For the image processing part, the image is

scaled down to 320x240 and is cropped afterwards to

a resolution of 320x108 (mainly sky and lower part

of the road are removed). The algorithm is running

on a Core2-Duo PC (each core running at 1.86 GHz),

that is stored in the side-bag of the motorcycle. The

average computation time at day is 28 ms and 6 ms at

night (almost dark). A display is connected to the PC

which shows the full rear-view image plus the indica-

tion in the upper image part (s. Fig. 3).

Figure 3: Overlay of triangle icon to indicate approaching

vehicle to the rider.

3.1 Recording Set-up

For recording, the prototype and an additional car

were equipped with GPS-sensors to get ground truth

data for relative speed and distance between motorcy-

cle and approaching vehicle. The GPS-data has been

synchronized with the video stream, which has been

captured frame-wise (uncompressed) by the PlaySta-

tion Eye camera.

The test-rides include high-speed conditions (car

is overtaking with up to 200 km/h while the motorcy-

cle is at 100 km/h), bad-weather and night conditions.

The overall recording times were 26 min of maneu-

vers with approaching cars and 1 hour 45 min with-

out any car in the video stream as well as sequences

where the motorcycle is standing still.

3.2 Data Evaluation

The ROC curves below (Receiver Operator Curve)

show the correct warnings (true positive rate, TPR)

against the false warnings (false positives per hour,

FP/h) for all recorded conditions. The TPR is event

based, which means that as soon as the algorithm de-

tects an approaching vehicle within a positive labeled

sequence, the detection is correct. Each positive la-

beled sequence has a ﬁxed length of 15s (time until

the approaching vehicle reaches the motorcycle). Fi-

nally, the TPR is the ratio between the number of cor-

rectly detected vehicles divided by the total number of

positive labeled sequences. The FP/h is estimated by

counting false warning events, i.e. consecutiveframes

of false warnings are interpreted as single event.

The ROC curves are estimated by increasing the

threshold t

(cf. Equation 10) by 0.01 for a range of

= [0.0 ... 10.0] (see Fig. 4). It can be directly per-

ceived, that the red ROC-curve is a bit jagged instead

of monotonically decreasing. This effect is because of

clustering consecutivefalse positives to a single event.

In special situations, the system still returns false

VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications

478

0 40 60 80 100 120

20%

40%

60%

80%

100%

true positive rate

false positives per hour

optimal working point

Figure 4: ROC curves of the system for values of t

ranging

from 0.0 to 10.0. The red curve represents the performance

of the system for recordings including the stroboscopic ef-

fect. The green curve shows the performance for recordings

without stroboscopic effect. The two arrows indicate the

same optimal working point for each ROC curve.

100

120

60 80

distance [m]

relative speed [km/h]

high speed

night

rain

approach

Figure 5: Scatter plot of detection distance (distance where

the vehicle has been detected) against relative speed (speed

differences between motorcycle and approaching vehicle)

for the optimal working point chosen from the ROC curves

above.

warnings (see red ROC curve in Fig. 4). Those are

due to the so called stroboscopic effect which hu-

mans also experience, e.g. on television (tyres of ve-

hicles are moving backwards while they are driving

forwards). Objects seem to move in different direc-

tions as in the real world, because of temporal alias-

ing effects caused by periodic motion at a rate close

to the frame rate of the camera. The proposed system

interprets this motion as true motion and warns the

rider in case that the motion is upscaling. The record-

ings contain such effects at an approximate speed of

120 km/h of the motorcycle while it passes a speciﬁc

bridge railing with periodic pattern. These bridges ap-

pear several times in the recordings, also because the

route has been driven multiple times.

If sequences including the stroboscopic effect are

removed from the data, the algorithm has an optimal

working range for values of t

= [1.7 ... 1.74], i.e.

TPR = 100% and FP/h = 0 (see green ROC curve).

The remaining recording times are 25 min of maneu-

vers with approaching cars and 1 hour 37 min without

any car in the video stream. Choosing the same work-

ing point for the red ROC-curve gives TPR = 100%

and FP/h = 32.

For the optimal working point in the ROC curve,

which corresponds to t

= 1.7, an additional scatter

plot is drawn (see Fig. 5). It depicts the distance and

relative speed between vehicle and motorcycle when

the vehicle has been detected for the ﬁrst time. Each

dot in the plot represents one driven maneuver. The

scatter plot shows four types of maneuvers: overtak-

ing at high speed (up to 100 km/h relative speed),

overtakingat night, overtakingin rainy conditionsand

approaching (car approaches on same lane as motor-

cycle).

For each maneuver,a line has been ﬁt into the data

set (solid line) with standard deviation (dashed lines).

As can be seen, the detection distances increase with

higher relative speed. Rainy conditions obviously do

not signiﬁcantly worsen the performance of the sys-

tem. This is due to the fact that the lens kept clean for

the whole ride because of the airstream. Surprisingly,

the detection distance at night is nearly constant at

approximately 60 m. The contrast at night around the

vehicle spotlights and the light cone in front of the ve-

hicle allow a very good motion estimation in the im-

ages. As soon as the vehicle is at a certain distance, so

that pixel movement is measurable, the algorithm is

able to immediately identify the moving front-lights.

4 CONCLUSIONS AND

OUTLOOK

Supporting the rider in observing blind spots and im-

proving the surround view compared to mirrors is an

important task to reduce accidents involving motorcy-

cles.

In this paper, a very robust and simple monocu-

lar approaching vehicle detection has been presented,

so that the rider is aware of other trafﬁc participants

behind or in blind spots. The system presented in

this paper relies on pixel motion information only and

hence is independent of the object appearance.

Extensive tests have been carried out under differ-

ent conditions including bad weather and rain. The

MonocularRearApproachIndicatorforMotorcycles

479

sensing distances span a range from 20 m at low rela-

tive speed (20 km/h relative speed) up to 60 m at night

conditions. The quantitative results are very promis-

ing so that the presented approach provides a cheap

and easy to implement support feature for motorcy-

cles.

Next steps will be undertaken to tackle the prob-

lem of the stroboscopic effect. Additionally, the view-

ing conditions concerning display size, display reso-

lution and camera ﬁeld of view will be optimized to

further increase the riding comfort.

ACKNOWLEDGEMENTS

The authors thank Achim Bendig for the produc-

tive teamwork and setting up the prototype. Spe-

cial thanks also to Nils Einecke for fruitful discus-

sions, and, last but not least, manythanks to Kazuyuki

Maruyama and Mutsumi Katayama for supporting the

project.

REFERENCES

Audi (2013). Audi side assist - innovative driver assis-

tance system. http://www.gizmag.com/audi-digital-

rear-view-mirror-production/23681/.

Bosch (2012). Side view assist. http://www.

bosch-automotivetechnology.com/en/de/homepage/

homepage

1.html.

Bouguet, J.-Y. (2000). Pyramidal implementation of

the lucas kanade feature tracker description of

the algorithm. http://robots.stanford.edu/cs223b04/

algo

tracking.pdf.

Fischler, M. A. and Bolles, R. C. (1981). Random sample

consensus: A paradigm for model ﬁtting with appli-

cations to image analysis and automated cartography.

Communications of the ACM.

Hella (2012). Driver assistance systems. http://www.

hella.com/hella-com/502.html?rdeLocaleAttr=en.

Klappstein, J., Stein, F., and Franke, U. (2006). Monocular

motion detection using spatial constraints in a uniﬁed

manner. IEEE Intelligent Vehicles Symposium.

Ma, Y., Soatto, S., Kosecka, J., and Sastry, S. S. (2004). An

Invitation to 3-D Vision. Springer-Verlag, New York,

2nd edition.

Mazda (2011). Mazda’s rear vehicle monitoring sys-

tem to receive euro ncap advanced award. http://

www.mazda.com/publicity/release/2011/201108/

110825a.html.

Mueller, D., Meuter, M., and Park, S.-B. (2008). Motion

segmentation using interest points. IEEE Intelligent

Vehicles Symposium.

Nissan (2012). Multi-sensing system with rear camera.

http://www.nissan-global.com/EN/TECHNOLOGY/

OVERVIEW/rear

camera.html.

OpenCV (2013). Open source computer vision library.

http://opencv.willowgarage.com/wiki/.

Quick, D. (2012). Audi’s digital rear-view mirror

moves from racetrack to r8 e-tron production vehi-

cle. http://www.gizmag.com/audi-digital-rear-view-

mirror-production/23681/.

Rabe, C., Franke, U., and Gehrig, S. (2007). Fast detec-

tion of moving objects in complex scenarios. IEEE

Intelligent Vehicles Symposium.

Scaramuzza, D. and Siegwart, R. (2008). Monocular om-

nidirectional visual odometry for outdoor ground ve-

hicles. Computer Vision Systems, Springer Lecture

Notes in Computer Science.

Schlipsing, M., Salmen, J., Lattke, B., Schroeter, K. G., and

Winner, H. (2012). Roll angle estimation for motorcy-

cles: Comparing video and inertial sensor approaches.

IEEE Intelligent Vehicles Symposium.

Shewchuk, J. R. (1996). Triangle: Engineering a 2d qual-

ity mesh generator and delaunay triangulator. Applied

Computational Geometry: Towards Geometric Engi-

neering.

Shewchuk, J. R. (2002). Delaunay reﬁnement algorithms

for triangular mesh generation. Computational Ge-

ometry: Theory and Applications, 22:1–3.

Stein, G., Mano, O., and Shashua, A. (2003). Vision-based

acc with a single camera: bounds on range and range

rate accuracy. IEEE Intelligent Vehicles Symposium.

Stierlin, S. and Dietmayer, K. (2012). Scale change and ttc

ﬁlter for longitudinal vehicle control based on monoc-

ular video. IEEE Intelligent Transportation Systems.

VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications

480