Improved algorithm for road region segmentation based on sequential monte carlo estimation

Natarajan Meghanathan et al. (Eds) : WiMONe, NCS, SPM, CSEIT - 2014
pp. 171–183, 2014. © CS & IT-CSCP 2014 DOI : 10.5121/csit.2014.41213
IMPROVED ALGORITHM FOR ROAD
REGION SEGMENTATION BASED ON
SEQUENTIAL MONTE-CARLO
ESTIMATION
Zdenek Prochazka
Department of Information Engineering,
National Institute of Technology, Oita College, Oita, Japan
zdenek_p@oita-ct.ac.jp
ABSTRACT
In recent years, many researchers and car makers put a lot of intensive effort into development
of autonomous driving systems. Since visual information is the main modality used by human
driver, a camera mounted on moving platform is very important kind of sensor, and various
computer vision algorithms to handle vehicle surrounding situation are under intensive
research. Our final goal is to develop a vision based lane detection system with ability to
handle various types of road shapes, working on both structured and unstructured roads, ideally
under presence of shadows. This paper presents a modified road region segmentation algorithm
based on sequential Monte-Carlo estimation. Detailed description of the algorithm is given,
and evaluation results show that the proposed algorithm outperforms the segmentation
algorithm developed as a part of our previous work, as well as an conventional algorithm based
on colour histogram.
KEYWORDS
Lane Recognition, Road Region Segmentation, Sequential Monte-Carlo Estimation
1. INTRODUCTION
In recent years, many car makers put a lot of intensive effort into development of autonomous
driving systems, and published their plans to start production of cars with self-driving ability
within next few years.
Autonomous driving systems exploit information from various sensors, and interpret sensory
information to identify appropriate navigation paths, as well as obstacles and relevant signage.
Among various sensors, a camera is an indispensable one, because real human driver rely mainly
on visual information, and transport infrastructure is most suited to the human vision system.
The development of computer vision techniques for autonomous driving is a very challenging
task, and intensive research on this topic is carried out. Our work deals with an issue of computer
vision based lane detection.
Several lane detection methods based on detection of road boundaries were already developed in
the past [1–5]. Detection of road boundaries is well suited for roads with clear surface marking,

172 Computer Science & Information Technology (CS & IT)
however it can easily fail if road boundaries are not clear, or if strong edges caused by non-road
objects are present in image. It is also not obvious, how to handle complicated road shapes,
splitting and merging of traffic line, complicated road marking, or sudden changes in road width.
To overcome the limitations of the boundary based approach, techniques for segmentation of road
as a region are intensively studied. A common issue with the road region segmentation is how to
build a classifier to make decision whether a image pixel belongs to the road surface or not. There
are two basic approaches to this issue. The first one is adopting an algorithm from field of
machine learning. Methods based on neural networks [6] or SVM [7,8] are reported. Since the
training of the classifier is generally an off-line process, the main difficulty with such kind of
methods is how to bring adaptability to new data which appear during detection stage and were
not part of the original training set.
The second approach are methods based on idea of statistical decision. Statistical distribution of
road features is modelled as parametric or non-parametric probability distribution function (pdf),
and the pdf is used to decide whether a given image pixel is part of road or not. Methods using
simple Gaussian colour model [9], Gaussian mixture model [10], histogram of colour components
[11], histogram of illumination invariant features [12] have been reported. The main difficulty
with the methods based on statistical decision is the fact, that it is not obvious how to estimate pdf.
To estimate pdf for each incoming frame, we need to some guess where the road is, which can
easily pose a chicken and egg problem. In many cases, some simple heuristics or ad-hoc solutions
are adopted. Furthermore, although the road segmentation should be performed over an sequential
image, many of the developed methods study segmentation only from single frame of road image,
and behaviour for image sequence is not studied enough.
In our work, we have focused to approach of statistical decision, and attempt to solve estimation
of pdf for sequential image in more systematic way. In our previous report, we have already
proposed a road segmentation algorithm based on sequential Monte-Carlo (SMC) estimation [13],
and applied it to region-based pathway estimation [14]. Although this segmentation method yields
better results than the conventional one, and it is quite robust to parameter setting, it still doesn’t
produce optimal result under certain circumstances. This paper shows a limitation of the original
algorithm and presents a new algorithm with improved segmentation ability. The proposed
algorithm is based upon a sound theoretical framework, and evaluation results for various kinds
of image data are shown to demonstrate effectiveness of our proposal.
Although an issue of shadows is not discussed in this paper, it should be mentioned that the
proposed method can be easily adapted to other type of features, which are claimed to be less
sensitive to illumination [10,12].
This document describes, and is written to conform to, author guidelines for the journals of
AIRCC series. It is prepared in Microsoft Word as a .doc document. Although other means of
preparation are acceptable, final, camera-ready versions must conform to this layout. Microsoft
Word terminology is used where appropriate in this document. Although formatting instructions
may often appear daunting, the simplest approach is to use this template and insert headings and
text into it as appropriate.
2. SMC ESTIMATION EASED ROAD SEGMENTATION ALGORITHM
This section at first recapitulates the original SMC based road region segmentation algorithm, and
discuss its limitations. After that a new algorithm with improved ability is proposed and described
in details.

Computer Science & Information Technology (CS & IT) 173
2.1. Former Algorithm
Let f be a vector of road features. In this work, a five dimensional feature vector f =
is used, where 1f and 2f corresponds to pixel coordinates x and y respectively,
and 3f , 4f , 5f corresponds to colour components bgr ,, . Although bgr ,, were adopted, the
algorithm is generally not limited to these features. For convenience we define also vectors
TT
ffyx ),(),( 21==x and TT
fffbgr ),,(),,( 543==c as shorthand for vector of coordinates
and colour components respectively. With this notation, the feature vector can be expressed also
as TTT
),( cxf = .
Figure 1. Original Algorithm
The goal is to estimate pdf of road features )( fp , and use the estimated pdf to classify road and
non-road pixels. Since the road can take various shapes, )( fp is expected to be a non-Gaussian
distribution at least with respect to the x . This suggests that we need an estimation method which
can handle general type of distribution. Since the road usually looks very similar between two
successive frames, and overall visual appearance changes relatively slowly compared to frame
rate, temporal continuity can be involved into the pdf estimation, and it should be possible to
model a visual changes of road as a stochastic process.
Taking the above mentioned into account, we have focused to SMC (Sequential Monte-Carlo)
estimation as a basic tool for estimation of the p ( f ). The original segmentation algorithm was
inspired by CONDENSATION algorithm [15]. The pdf is expressed in discrete form as a set of
weighted samples S ,
{ }),(),,(),,( 2211 NN wwwS fff Λ= (1)

where if is a sample of the feature vector f and iw is an importance weight assigned to the
sample. Estimation is performed by repetitive update of the S according to the algorithm shown
in Figure 1. The superscript k stands for discrete time here. The ν in the eq.(2) means a vector
of Gaussian random numbers with zero mean and variances 51, σσ Λ . The )|( iip cz in eq.(3) is
an observation pdf defined by eq.(5).
(a) Original Image (b) Samples (c) Segmentation Result
Figure 2: Example of Segmentation Result






−Σ−−= −
)()(
2
1
exp)|( 1
iim
T
iiiip zczccz (5)
The symbol mΣ here means a diagonal matrix of variances, controlling width of exponential
function for each colour component, and iz means a colour component observed at position
obtained from eq.(2). Simply said, the )|( iip cz is a measure of similarity between colour
components ic predicted by eq.(2), and colour components observed at predicted positions. The
calculated similarity is than assigned as an importance weight.
The above described algorithm estimates the pdf in a form of weighted samples. To evaluate pdf
for an arbitrary pixel, we perform a smoothing of the estimated pdf, in a way similar to non-
parametric density estimation, using kernel function φ . Hence, the pdf )( fp for an arbitrary
pixel values f is calculated as
∑=
=
N
i
iiwp
1
),()( fff φ (6)






−Σ−−= −
)()(
2
1
exp),( 1
is
T
ii ffffff βφ
where the symbol sΣ is a diagonal matrix of parameters, controlling window width for each
component of f , and the β is a normalising constant. Segmentation of the road region is
performed by evaluation of )( fp for each pixel, followed by thresholding of obtained values to
decide road and non-road pixels.
We have already reported that the above described segmentation method is quite robust to
parameter tuning and yield better results than conventional segmentation methods like region
growing or simple Gaussian colour model [16]. However there are still issues to be solved. Let’s
look at Figure 2 to for explanation.

The image shown in Figure 2(a) is a synthetic image, and the road region can be segmented very
easily here.The Figure 2(b) shows a spatial distribution of samples if , after the method described
above has been applied. As can be seen, the samples are not distributed uniformly, and only few
samples appear over left and right side of low part of the road region. As a consequence, these
parts are not segmented properly, as is shown in Figure 2(c). Such improper distribution of
samples can cause serious problems in proper segmentation of roads with multiple traffic lines.
There is another issue, which can’t be simply shown by static image. Since the estimation of pdf
of the road features is treated as an stochastic process, certain noise has to be introduced into the
model. Due to this stochastic nature, the positions of samples change between subsequent frames,
and segmented parts of road covered only by few of samples are very sensitive to such changes.
Due to this, rapid changes in shape of the segmented road appear from frame to frame, although
the road doesn’t changes significantly in the image. We refer these quick temporal changes of
segmentation result as fluctuations in the following. It is clearly an unwelcome phenomenon
which have to be eliminated.
After careful revision of the original algorithm, we have proposed a new algorithm, which
attempts to solve the above mentioned issues.
2.2. Proposed Algorithm
The limitations of the original algorithm are caused by improper distribution of the samples
during the estimation process. To obtain proper segmentation result, samples should be spread
uniformly over the road region, and the sampling method described in step 4 of the Figure 1 can’t
guarantee this property. In the original algorithm, the only way to spread samples more widely
over the road region is to increase amount of transition noise added to spatial components.
However this cause larger fluctuations of the segmented region, therefore limitations of the
original algorithm cannot be simply solved by parameter tuning.
Figure 3: Assignment Between Subsequent Samples
Our idea to solve discussed issues is based on the principle of importance sampling [17]. In this
framework, it is supposed that samples can be easily drawn from some probability distribution
function ),|(
)()1()( k
ii
k
iq zff
−
, which is called importance density. For samples generated from
the importance density, an update of the sample weights is done according to the following form.

),|(
)|()|(
)()1()(
)1()()1()(
)1()(
kk
i
k
i
k
i
k
i
k
i
k
k
i
k
i
q
pp
ww
zff
fffz
−
−−
−
∝ (7)
Although principle of importance density provides sound theoretical framework, the specific
forms of )|( )1()( −k
i
k
p fz , )|( )1()( −k
i
k
ip ff and ),|(
)()1()( k
ii
k
iq zff
−
are not obvious, and
strongly depends on particular application.
In our case, we have chosen ),|(
)()1()( k
ii
k
iq zff
−
as a uniform distribution with respect to spatial
components x , where the range of x is limited to region segmented at the previous time step.
Simply said, spatial components of the generated samples ix ),,1( Ni Λ= are uniformly drawn
from the region obtained at 1−k .
To evaluate )|( )1()( −k
i
k
ip ff , we need to establish relations between old samples and new
samples generated according to importance density. This is not straightforward, and our idea is to
perform assignment in a way described in Figure 3 for single variate case. Let the 41 ,, xx Λ be
old samples, and let the intervals highlighted by light grey be segmented parts of x axis. We draw
new samples from the light grey intervals randomly, according to a uniform distribution. New
samples are shown by diamond and the alphabet subscript here. For each new sample
ix ),,,( dcbai = we assign one of the old samples jx )4,3,2,1( =j , using the weighted kernel
function φ , already mentioned in eq.(6), as an assignment criteria.
)),((maxarg
4,,1
jixj
j
xxwl φ
Λ=
= (8)
Here, xφ stands for univariate Gaussian kernel function. The l obtained by eq.(8) is than used as
a subscript of assigned old sample. For the situation in Figure 3, the resulting assignment is
),( 1xxa , ),( 3xxb , ),( 2xxc , ),( 4xxd .
We have described the basic idea of sample assignment for simple univariate case. Next we
explain, how to extend this idea to the our case of f . Since our importance density is defined in
spatial subdomain x , we perform assignment of samples with respect to x . Therefore, after the
new samples )(k
ix are generated according to importance density, we assign to each generated
sample an old one, for which the ( ))1()()1(
, −− k
j
k
i
k
jw xxxφ is maximal. The symbol xφ here stands
for kernel function of the same form as in eq.(6), however applied only to spatial components
.
By the above described procedure, we can generate spatial part )(k
ix of the new samples, and
perform assignment between new and old samples. However, we said nothing about the colour
part )(k
ic so far. To generate colour part of the new samples, and build up complete samples
)(k
if , we simply apply transition equation to colour part of assigned old sample.
ic
k
l
k
i νcc += − )1()(
(9)
Here the ic ν means a colour part of ν from the eq.(2), and the subscript l means a label of an
assigned old sample.

After the new sample set
)(k
if is completed, we evaluate transition pdf in weight update equation
as
)|()|( )1()()1()( −−
= k
l
k
i
k
l
k
ip ffff φ (10)
where φ has a same form as in eq.(6), andl is a label of sample assigned to the i -th one.
The last component needed for weight update calculation in eq.(7) is the observation pdf
)|( )1()( −k
i
k
p fz , and it has the same form as was already shown by eq.(5).
Figure 4. Proposed Algorithm
The proposed algorithm is summarised in Figure 4.In the weight update eq.(12) of the step 8, the
value of importance density q is omitted, since we draw samples according to uniform
distribution.
At the end of this section, we show a simple example of the segmentation result obtained by the
proposed algorithm. The result for the same data as in Figure 2 is shown in Figure 5. As can be
seen, samples are spread over all road region and there are no missing parts of segmented road
region.
Figure 5: Example of Segmentation Result

In the next section, we present more detailed evaluation of the proposed algorithm for various
kinds of image data.
3. EVALUATION OF THE PROPOSED ALGORITHM
The evaluation of the proposed algorithm have been performed on the following kinds of image
data.
A: Sequences of Still Images
Approximately 200 images, capturing frontal roads viewed from car, were randomly
collected. Half of them are structured roads, second half are unstructured and rural roads.
For the each image, a sequential image was created by repeating of the same still image.
Created sequential images shows snapshots of real road, but since there are no temporal
changes between subsequent frames, these data can be used to assess fluctuations of the
resulting segmentation.
B: Synthetic Image Sequence
Sequence of rendered CG images, already used in Figure 2 and Figure 5, was created.
The sequence simulates a vehicle moving by 60 km/h on road with 4 curves (R=150m,
100m, 80m, 50m). Vehicle steering is simulated by changes in lateral position and pitch
angle. Since the sequence is composed of road images with almost homogeneous colour,
the perfect segmentation of road region can be done very easily, and this sequence can be
used to assess quality of segmentation.
C: Real Image Sequences
Real image sequences, already mentioned in our report [14], with representative frames
shown in Figure 6. These sequences are real data, therefore they can be used to assess
overall ability of the algorithm.
For all above mentioned data, ideally segmented road regions were prepared as ground truth. The
segmentation was done manually for data A and C, and automatically during CG rendering for
data B.
As a quantitative measure of the segmentation ability, we adopted Jaccard index J , defined by
eq.(14).
||
||
gs
gs
RR
RR
J
∪
∩
= (14)
Here, sR is a set of road pixels obtained by segmentation algorithm, and gR is a set of road pixels
for ground truth. Symbol | | means a set size here. Jaccard index J takes value 1 if both sets sR
and gR are identical.
Besides the SMC based algorithms described in this paper, we implemented another segmentation
algorithm based on seeded region growing (SRG). Overall configuration is shown in Figure 7.

Figure 6: Real Image Sequences
SRG based algorithm use normalised histogram of features as a merging criteria (pixel is merged
if histogram value for pixel features is greater than certain threshold). The histogram of features is
calculated from region placed over low part of road image. This follows the same strategy as I
reports [11] and [12]. Placement of seed points follows suggestions given by report [12].
SRG based algorithm, original SMC based algorithm and proposed SMC based algorithm were
evaluated for data A, B, C. Before evaluation, all necessary parameters were once tuned to
maximise average value of Jaccard index eq.(14). After tuning, parameter values were fixed, and
same fixed values were used for all tested data.
Figure 7: Configuration of SRG Based Reference Algorithm
3.1. Evaluation for Sequences of Still Images
Three mentioned algorithms, were applied to data A, and Jaccard index eq.(14) was calculated for
segmentation results of each particular sequence frame by frame. Next, average Jµ and standard
deviation Jσ of Jaccard index was calculated separately for each sequence, and these values were
used to form histograms of Jµ and Jσ respectively. The resulting histograms are shown in
Figure 8. Since data A have no temporal changes, results by SRG based algorithm contains no
fluctuations, and evaluation of Jσ has no meaning for data A. Therefore, only result for SMC
based algorithm is shown in Figure 8(b). Ideal values of Jµ and Jσ are 1=Jµ , 0=Jσ , where
higher Jµ indicates better segmentation result, and lower Jσ indicates less amount of
fluctuations. Figure 8(a) shows that peak of histogram for the proposed algorithm is closer to

ideal value 1=Jµ , and (b) of the same figure shows that peak of histogram for the proposed
algorithm is closer to ideal value 0=Jσ . It means, that for majority of the data A, the proposed
algorithm yields segmentation results closer to ground truth, while fluctuations of segmented
regions are reduced. The result for SRG based algorithm is surprisingly wrong. The values of
average of Jaccard index are distributed in wide range, which shows that the SRG based
algorithm is not robust to parameter setting.
(a) Histogram of Means Jµ (b) Histogram of Standard Deviations Jσ
Figure 8: Histogram of Jaccard Index for Sequences Still Image
3.2. Evaluation for Synthetic Image Sequence
Three mentioned algorithms were applied to data B, and Jaccard index eq.(14) was calculated for
each segmented frame of image sequence. Graphs of the Jaccard index are shown in Figure 9. It
can be seen, that proposed algorithm yields similar result as SRG based algorithm, where result in
Figure 9(a) is more noisy than Figure 9(c). This is caused by fact, that SRG based algorithm
segments each frame as an independent still frame and there is no mechanism for temporal
continuity. The result for original algorithm in Figure 9(b) contains large fluctuations, and
magnitude of fluctuations is significantly reduced by the proposed algorithm.
(a) SRG Based (b) SMC Based Original (c) SMC Based Proposed
Figure 9: Jaccard Index for Synthetic Image Sequence
3.3. Evaluation for Real Image Sequences
Three mentioned algorithms were applied to data C, and Jaccard index eq.(14) was evaluated for
segmentation result of each tested sequence frame by frame. Graphs of the Jaccard index are
shown in Figure 10. The results show similar tendency as for data A and B. SRG algorithm
yields quite unstable results. The original SMC based algorithm shows high values of J for
sequences No.1-3, but fluctuations are observed. The results of original algorithm for sequences
No.4-6 shows lower values of J . Road in this images have multiple traffic lines, and due to the

issue discussed in subsection 2.1, other traffic lines are not segmented properly. The proposed
algorithm shows lower magnitude of fluctuations and overall shows high values of J . Several
decreases of J can be observed in sequence No.5. Let us now look more closely on segmentation
results fore such frames. Figure 11
(a) SRG Based (b) SMC Based Original (c) SMC Based Proposed
Figure 10: Jaccard Indexes for Real Image Sequence
shows result for frame 125 of sequence No.5. This is a situation immediately after another car has
passed an opposite traffic line. Since samples disappear from road parts occluded by passing car,
these parts are not segmented, and a few iterations of the algorithm are needed to spread samples
over the previously occluded road parts. Due to this the Jaccard index decreases for few frames,
because the segmented region differs from its ground truth.
The evaluation results for data A, B, C described above shows, that the proposed algorithm pose
an effective solution to the issues discussed in subsection 2.1.
Figure 11: Frame 125 of sequence No.5

4. CONCLUSION AND FURTHER WORK
This paper has presented a road segmentation algorithm based on sequential Monte-Carlo
estimation. The proposed algorithm is an extension of our previous work on topic of road region
segmentation, and the proposed algorithm solves issues remained in original SMC based
segmentation algorithm.
The key idea of the proposed algorithm lies in improvement of sampling method and consequent
calculation method of importance weights. Drawing samples according to distribution of
importance weights, which was adopted in the original algorithm, was not optimal for road region
segmentation, and more general approach based on principle of importance sampling has been
introduced. Particularly, uniform sampling from region segmented at previous time step was
introduced to generate new sample set, and assignment between subsequent samples was
introduced to evaluate transition probability density function required in weight update.
The proposed algorithm was evaluated on three types of image data, and evaluation results shows
superiority of the proposed algorithm over the original one, as well as over the algorithm based on
seeded region growing.
In the proposed algorithm, RGB colour components were used as features, however the proposed
algorithm can be easily adapted to different features too. Evaluation of the proposed algorithm
for features which are more resistant to illumination changes, and adaptation of the algorithm for
road images with hard-cast shadows is a main direction for future work.
REFERENCES
[1] Y. Wang, D. Shen, and E. Teoh, “Lane detection using catmullrom spline,” in IEEE International
Conference on Intelligent Vehicles, 1998, pp. 51–57.
[2] B. Ma, S. Lakshmanan, and A. O. Hero, “Pavement boundary detection via circular shape models,” in
Proceedings of the IEEE Intelligent Vehicles Symposium 2000, 2000, pp. 644–649.
[3] Y. Wang, E. Teoh, and D. Shen, “Lane detection and tracking using b-snake,” ImageVision and
Computing , vol. 22, no. 4, pp. 269–280, Apr. 2004.
[4] H. Xu, X. Wang, H. Huang, K. Wu, and Q. Fang, “A fast and stable lane detection method based on
b-spline curve,” in IEEE 10th International Conference on Computer Aided Industrial Design
Conceptual Design. Ieee, 2009, pp. 1036–1040.
[5] X. Miao, S. Li, and H. Shen, “On-board lane detection system for intelligent vehicle based on mono-
cular vision,” International Journal on Smart Sensing and Intelligent Systems , vol. 5, no. 4, pp. 957–
972, Dec. 2012
[6] M. Foedisch and A. Takeuchi, “Adaptive real-time road detection using neural networks,” in Proc.
7th Int. Conf. on Intelligent Transportation Systems, Washington D.C , 2004.
[7] S. Zhou, J. Gong, G. Xiong, H. Chen, and K. Iagnemma, “Road detection using support vector
machine based on online learning and evaluation,” in Intelligent Vehicles Symposium 2010, Jun.
2010, pp. 256–261.
[8] E. Shang, X. An, L. Ye, M. Shi, and H. Xue, “Unstructured road detection based on hybrid features,”
in Proceedings of the 2012 2nd International Conference on Computer and Information Applications
(ICCIA 2012), Dec. 2012, pp. 926–929.
[9] J. Crisman and C. Thorpe, “Scarf: A color vision system that tracks roads and intersections,” IEEE
Trans. on Robotics and Automation, vol. 9, no. 1, pp. 49 – 58, February 1993.
[10] O. Ramstrom and H. Christensen, “A method for following of unmarked roads,” in Proceedings of
Intelligent Vehicles Symposium 2005. IEEE, 2005, pp. 650–655.
[11] S. K. O. Changbeom, S. Jongin, “Illumination robust road detection using geometric information,” in
15th International IEEE Conference on Intelligent Transportation Systems (ITSC2012), 2012, pp.
1566–1571.

[12] J. Alvarez-Mozos, A. Lopez, and R. Baldrich, “Illuminant-invariant model-based road segmentation,”
in IEEE Intelligent Vehicles Symposium 2008 , Eindhoven, Netherlands, Jun. 2008, pp. 1175 –1180.
[13] Z. Prochazka, “Road region segmentation based on sequential monte-carlo estimation,” in
Proceedings of ICARCV 2008 , Dec. 2008, pp. 1305–1310.
[14] Z.Prochazka, “Pathway estimation for vision based road following suitable for unstructured roads,” in
Proceedings of ICARCV 2012 , Dec. 2012, pp. 1374–1379.
[15] M. A. Isard, “Visual motion analysis by probabilistic propagation of conditional density,” Ph.D.
dissertation, University of Oxford, Sep 1998.
[16] Z. Prochazka, “Road tracking method suitable for both unstructured and structured roads,”
International Journal of Advanced Robotic Systems , vol. Vol. 10, no. No. 158, pp. 1–10, 2013.
[17] A. Doucet, S. J. Godsill, and C. Andrieu, “On sequential monte carlo sampling methods for bayesian
filtering,” Statistics and Computing, vol. 10, no. 3, pp. 197–208, 2000.
AUTHORS
Zdeněk Procházka received M.Sc degree in radio-electronics in 1994 from the
Faculty of Electrical Engineering, Czech Technical University in Prague, Czech
Republic. In 2000 received Ph.D degree from University of Electro-Communications
in Tokyo, Japan. From 2007 associated professor of National Institute of Technology,
Oita College in Oita, Japan.

Improved algorithm for road region segmentation based on sequential monte carlo estimation

More Related Content

What's hot (19)

Viewers also liked (20)

Similar to Improved algorithm for road region segmentation based on sequential monte carlo estimation (20)

Recently uploaded (20)

Improved algorithm for road region segmentation based on sequential monte carlo estimation