Real-time Stereo Vision System at Tunnel

Yuquan Xu

, Seiichi Mita

, Hossein Tehrani

and Kazuhisa Ishimaru

Research Center for Smart Vehicles, Toyota Technological Institute, 2-12-1 Hisakata, Tempaku, Nagoya, Japan

Driving Assist & Safety Eng. Div. 1, DENSO Corporation, 1-1, Showa cho, Kariya, Japan

Research & Development Dept. 2, Nippon Soken Inc., Nishio, Aichi, Japan

Keywords:

Stereo Vision, Image Deblurring, Optical Flow, Cepstrum.

Abstract:

Although stereo vision has made great progress in recent years, there are limited works which estimate the

disparity for challenging scenes such as tunnel scenes. In such scenes, owing to the low light conditions and

fast camera movement, the images are severely degraded by motion blur. These degraded images limit the

performance of the standard stereo vision algorithms. To address this issue, in this paper, we combine the

stereo vision with the image deblurring algorithms to improve the disparity result. The proposed algorithm

consists of three phases: the PSF estimation phase; the image restoration phase; and the stereo vision phase.

In the PSF estimation phase, we introduce three methods to estimate the blur kernel, which are optical ﬂow

based algorithm, cepstrum base algorithm and simple constant kernel algorithm, respectively. In the image

restoration phase, we propose a fast non-blind image deblurring algorithm to recover the latent image. In the

last phase, we propose a multi-scale multi-path Viterbi algorithm to compute the disparity given the deblurred

images. The advantages of the proposed algorithm are demonstrated by the experiments with data sequences

acquired in the tunnel.

1 INTRODUCTION

In recent years, signiﬁcant attention has been paid

to the development of autonomous vehicles and ad-

vanced driver assistance systems (ADAS). Stereo vi-

sion is an important research problem for ADAS ap-

plications and has received a great deal of attention

over the past decade. However, there is little work

considering the stereo vision in some challenge envi-

ronments, such as tunnels or low lightening condition.

For the ADAS applications, the stereo camera

pair, mounted on the vehicle, is used to estimate

the environment’s depth information from images ac-

quired by the stereo pair. However, in certain environ-

ments, such as the tunnel, the low illumination and

the fast camera movement, result in the blurring of

the stereo pair images. The blurry images reduce the

quality of stereo’s depth estimation. The motion blur

is caused by the relative movement between the cam-

era and the object during the camera’s exposure time.

The blur can be attributed to three tunnel-speciﬁc fac-

tors in ADAS applications. 1) The speed of the car,

which is typically very high, since there are no trafﬁc

lights or sharp turns in the tunnel. 2) The distance be-

tween the camera and the object. The tunnel area is

very limited and the walls and roof of the tunnel are

near to the vehicle. 3) The length of exposure time,

which is high to account for the low illumination of

the tunnel. Stereo vision can be viewed as a matching

problem and for blurry scenes there may be multiple

matching pixels across the left and right image pair.

This phenomenon violates the basic assumption of the

stereo vision framework. Fig. 1 shows the compari-

son results of a tunnel example using the proposed

algorithm and other state-of-the-art real-time stereo

vision algorithms, Semi-Global Block-Matching Al-

gorithm (SGM) (Hirschm

uller, 2008) and Multi-paths

Viterbi (MPV) (Long et al., 2014b). In Fig. 1, we

can see that the input image is highly blurred espe-

cially on the left wall, and the previous stereo vision

approaches can not reliably estimate the disparity for

such images. However, the proposed algorithm can

produce high quality disparity results, even for the

challenging scenes.

A straight forward approach to address this issue

involves the deblurring of the degraded input stereo

pairs to improve the disparity results. Motion deblur-

ring is an important research problem in low-level vi-

sion research. There are two research sub-problems

that are addressed in this research. The ﬁrst problem

involves the estimation of the Point Spread Function

402

Xu Y., Mita S., Tehrani H. and Ishimaru K.

Real-time Stereo Vision System at Tunnel.

DOI: 10.5220/0006112304020409

In Proceedings of the 12th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2017), pages 402-409

ISBN: 978-989-758-227-1

(a) (b)

Figure 1: Visual comparison of our method to SGM

(Hirschm

uller, 2008) and MPV (Long et al., 2014b) on an

example captured in the tunnel. (a) is the left image of

the stereo pair. (b) MPV (Long et al., 2014b). (c) SGM

(Hirschm

uller, 2008). (d) The proposed method. In the dis-

parity map, red represents near objects and blue represents

distant objects.

Figure 2: The ﬂowchart of the proposed algorithm.

(PSF) or so-called blur kernel which represents the

degradation process. The second problem involves

the recovery of the unknown latent image. Typi-

cally, most of the existing deblurring methods (Cho

and Lee, 2009; Xu and Jia, 2010) aim to deal with

the spatial-invariant blurs which are modeled as a de-

convolution problem. However, in the tunnel case,

the blur is spatially-variant. Furthermore, for ADAS

applications, the deblurring and disparity estimation

should be performed in real-time, that means the run-

ning time should be lower than 100ms. However,

other deblurring algorithms usually cost minutes to

deblur one image, which is unacceptable for real-time

application.

In this article, we introduce a real-time algorithm

to deal with the tunnel stereo vision problem. The

proposed algorithm consists of three phases, i.e., PSF

estimation phase, image restoration phase and stereo

vision phase. In the PSF estimation phase, the tradi-

tional iteration based framework’s computational cost

is reduced by utilizing three different algorithms to

directly estimate the spatially-variant blur kernel by

using optical ﬂow, cepstrum and a simple constant

kernel. After we estimate the PSF of the image, a fast

non-blind image deblurring algorithm is introduced to

remove the blur effect of the images. Last we propose

a multi-scale MPV algorithm to compute the disparity

result. Our proposed approach is real-time which is

signiﬁcantly important for practical applications. We

show the ﬂowchart of the proposed algorithm in Fig.

2 RELATED WORK

Stereo vision is one of the basic problems in com-

puter vision and has made tremendous progress in

last decades. Scharstein and Szeliski (Scharstein and

Szeliski, 2002) introduced a categorization scheme

for stereo algorithms, which classiﬁed the various

stereo algorithms into two classes including local and

global methods. Local algorithms usually contain

four steps: (1) matching cost computation; (2) cost

aggregation; (3) disparity computation; (4) the op-

tional post processing. The matching cost compu-

tation methods include standard window-based algo-

rithms, normalized cross correlation (NCC), Census

Transform algorithm (Zabih and Woodﬁll, 1994), and

recently CNN based cost (Zbontar and LeCun, 2014).

The cost aggregation algorithms contain unnormal-

ized box ﬁltering, bilateral ﬁlter and guided image ﬁl-

ter (He et al., 2013). On the other hand, the global

methods make explicit smoothness assumptions and

then solve an optimization problem from the match-

ing cost and omit the cost aggregation step. The op-

timization algorithms include dynamic programming

(Veksler, 2005), belief propagation (Sun et al., 2003),

Viterbi (Long et al., 2014a; Long et al., 2014b) and

graph cuts (Boykov et al., 2001). SGM (Hirschm

uller,

2008) is a hybrid of local and global method that ex-

pands the single-directional 1D scanline optimization

process to multidirectional 1D scan-line optimization.

Removing the blur effects of the images is a well-

studied problem. For spatial-invariant blind image

deblurring issue, Fergus et al. (Fergus et al., 2006)

ﬁrstly proposed a successful variational Bayesian

framework to estimate the complex motion blur ker-

nel. In (Cho and Lee, 2009), a fast deblurring method

was proposed by introducing a prediction step. The

similar step was used in (Xu and Jia, 2010) (Xu et al.,

2012). For spatial-variant blur issue, Joshi et al.

(Joshi et al., 2010) introduced inertial measurement

sensors to record the motion path of the camera to

recover the latent image. Then the projective motion-

blur model was introduced by (Tai et al., 2011) and

used in (Whyte et al., 2011) (Hirsch et al., 2011).

Zheng et al. (Zheng et al., 2013) proposed a forward

motion model which is more relevant to our work, but

Real-time Stereo Vision System at Tunnel

403

(a) (b)

Figure 3: (a) is the blurry input, (b) is the blur kernel in each

local area of the image.

their work is too time-consuming for practical appli-

cation. We refer the readers to the (Wang and Tao,

2014) for the recent progress of the image deblurring.

3 STEREO VISION ALGORITHM

In this section, we introduce a coarse-to-ﬁne frame-

work to estimate the disparity result based on the de-

blurred images. We adopt the structural similarity

(SSIM) to measure the matching cost between left and

right deblur images.In this coarse-to-ﬁne framework,

we generate the 3 layers Gaussian pyramid of the de-

blurred images, and each coarser layer is half size of

the ﬁner layer. Then we utilize the MPV (Long et al.,

2014a) method to estimate the disparity result from

the coarsest to the ﬁnest layer. The cost function of

proposed algorithm is:

E(D) = E

(D) +E

(D) (1)

Where D represents the disparity result, E

(D)

represents the SSIM cost of deblurring images and

(D) represents the TV constraints of disparity.

Figure 4: Viterbi path to ﬁnd optimum disparity.

The problem of stereo matching can be modeled

as ﬁnding the disparity map D that minimizes the en-

ergy function E(D). In this case, the Viterbi algorithm

can be used to search the optimum solution (Son and

Seiichi, 2006). We build a graph to represent the dis-

parities of the image pixels along one line, where each

node in this graph represents a disparity assigned to a

pixel and each edge represents a candidate disparity

change between two pixels in the same path. Viterbi

algorithm operates on individual rows and ﬁnd the op-

timum ”path” of nodes through disparity space from

one side of the image to the other as shown in Fig 4.

To ﬁnd the optimum disparity value, we solve follow-

ing cost function:

= arg min

V (t −1,d

t−1

) + SSIM(t, d)+ λ|d − d

t−1

(2)

where d

represents the disparity at pixel t, V (t −

1,d

t−1

) represents the energy of a node with pixel t −

1 with disparity d

t−1

, SSIM(t,d) represents the SSIM

cost with pixel t and disparity d and λ|d −d

t−1

| is the

smoothing penalty term.

To increase the robustness of the Viterbi algo-

rithm, four bi-directional (horizontal, vertical, and 2

diagonals) Viterbi paths are used to provide good cov-

erage of the 2D image. The ﬂowchart of the Multi-

path Viterbi algorithm is shown in Fig. 5.

With the results on the coarser layer, we only con-

sider a small candidate disparity range. This multi-

scale MPV algorithm is much faster than (Long et al.,

2014a) and more stable to get better results on some

textureless area.

Figure 5: The ﬂowchart of the Multi-path Viterbi algorithm.

4 PSF ESTIMATION

In this section, we estimate the motion kernel of the

tunnel images. For a given image with long exposure

time, the estimating the motion blur kernels are not

trivial. Fortunately in the tunnel case, the exposure

times are limited by the frame rate. Additionally, the

vehicle typically travels in a straight path in the tun-

nel. Therefore, the motion observed during the expo-

sure will be relatively smooth and the PSF in a small

local area can be approximated as the linear kernel, as

VISAPP 2017 - International Conference on Computer Vision Theory and Applications

404

Figure 6: The ﬁrst column is the original image. The second to last column are the disparity results of SGM (Hirschm

uller,

2008), ELAS (Geiger et al., 2011), MPV (Long et al., 2014b) and the proposed method respectively.

shown in Fig. 3. This approximation helps in the case

of the spatially-varying blur since estimating an arbi-

trary kernel for each individual pixel would be pro-

hibitively expensive. Based on this assumption, the

blur kernel at each pixel is a straight line which is de-

termined by two parameters, i.e., the length and the

angle. Then the blurry input can be formulated as:

b =

∑

[(ω

∗ l) ⊗ f

] + n (3)

where b is the blurry input, ω

is the window func-

tion to select a sub-region of the image, l is the latent

image we want to recover, i is the index of the sub-

region, ∗ denotes the element-wise multiplication, ⊗

denotes the convolution process, Z is normalization

factor of the window function, n is the additive noise,

represents the blur kernel of the i −th sub-region of

the image, respectively. For the PSF, we have:

f (x, y) =



1/L others

0 y = x tan θ, 0 ≤ x ≤ L cos θ

(4)

where x and y are pixel index and L is the length of the

kernel, θ is the angle of the kernel, respectively. Then

we introduce three algorithms to estimate this kind of

blur kernel for the blurry input.

4.1 Optical Flow

Since the motion blur is caused by the relative motion

of the camera and objects, it follows that the knowl-

edge of the actual camera motion between consecu-

tive image pairs that can provide signiﬁcant informa-

tion when performing image deblurring. We assume

that the camera moves nearly along a straight line be-

tween two successive frames and the optical ﬂow of

these two frames can represent the blur kernel for the

blurry input. To compute the optical ﬂow, we have the

cost function as Equ. (5)

E(w) = E

(w) + λE

(w) (5)

(a) (b) (c)

Figure 7: (a) is the blurry image and the blur kernel in its

right up. (b) is the cepstrum of blurry input. (c) is the value

alone the blur angle in the cepstrum.

where w is the ﬂow vector, λ is the smoothing param-

eter, E

(w) is the data term and E

(w) is smoothing

term, respectively. For E

(w) and E

(w) we have:

(w) =

(u + w) − I

(w)

∇I

(u + w) − ∇I

(w)

(6)

(w) =

∇w

(7)

where I

and I

are two successive frames, ∇ =

(∂

,∂

) represents the gradient of the image, α ∈ [0,1]

denotes a linear weight. The smoothing term is the

Total Variation (TV) constraint. We use the Suc-

cessive Over-Relaxation (SOR) method to solve Equ.

(5). Since the input image is blurry and we actually

don’t need the optical ﬂow for every pixel of the im-

age

, we downsample the input left and right image

to 1/4 size which can not only reduce the blurry effect

in the image but also reduce the computation time of

the algorithm. Then we upsample the optical ﬂow to

represent the blur kernel of the image.

4.2 Cepstrum

To deblur the image, we ﬁrst segment the blurry in-

put into several overlapped patches and assume that

in each sub-region the blur kernel is a straight line.

Actually we deblur the input image block by block, so

we only need an optical ﬂow for every small local area of

the input, not every pixel

Real-time Stereo Vision System at Tunnel

405

Based on this assumption, we can compute the cep-

strum of the image to identify the angle and length of

this kind of linear kernel. The cepstrum of an image

is deﬁned as (Rom, 1975):

C(g) = F

−1

(log(|F(g)|) (8)

where F and F

−1

denote the Fourier transform and

inverse Fourier transform , C(g) is the cepstrum of

the image g. In practice, the cepstrum of the image is

usually expressed as:

C(g) = F

−1

(log(|F(g)| + 1) (9)

This linear degrade process for each sub-patch is

described as Equ. (10)

= s

⊗ f + n (10)

where s

and s

denote the sub-patch of the blurred

and clear image.

In the case of ignoring the additive noise n, we can

formulate Equ. (10) to cepstrum domain as:

C(s

) = C(s

) +C( f ) (11)

where C(s

), C(s

) and C( f ) represent the cepstrum

of blurred image s

, s

and f . Therefore the cepstrum

of blurred image is modeled as the sum of the cep-

strum of the clear image and the cepstrum of a blur

kernel, which means the convolution operation be-

comes additive in cepstrum domain, thus the blur is

easy to detect.

Fig. 7 shows an example of how we use the cep-

strum to detect the blur kernel of the image. Fig. 7

(a) shows a blurry and noisy image. The blur ker-

nel has 20 pixels length and 30-degree angle and the

variance of the Gaussian noise is 0.03. We show the

cepstrum of the blurry input in Fig. 7 (b), in which we

can clearly see a straight line along the 30-degree. In

practice, we adopt the Radon transform to detect this

angle as the blur angle and then we rotate the cep-

strum to make this line to be horizontal. Fig. 7 (c)

shows the values on this line of the cepstrum and we

can see two negative peaks in this line. We compute

the distance between these two negative peaks and the

blur length is half of the distance. Finally, we can es-

timate the angle and length of the blur kernel. For a

blurry tunnel input, we segment it to some overlapped

cloacal areas and use the cepstrum to estimate the blur

kernel of this sub-region.

4.3 Proposed Method

In autonomous vehicle applications, the stereo vision

algorithm is required to generate the disparities in

real-time continuously and stably. Although we have

Table 1: Computation time.

Methods Time Environment

Optical ﬂow 9.8s Intel i7 4790 CPU

Cepstrum 7.8s Intel i7 4790 CPU

Proposed 1.4s Intel i7 4790 CPU

Proposed 96ms GTX TITAN X GPU

proposed two algorithms to estimate the blur kernel,

the computational time is not real-time or less than

100ms. Therefore, we further reduce the computa-

tional complexity of the deblurring process for the

stereo pairs. More speciﬁcally, we adopt a constant

kernel as an approximate motion kernel for all the

pixels of the tunnel images as shown in Fig. 8 (d).

In this constant kernel, we pre-deﬁne a point in the

image to represent the vanishing point along the trav-

eling direction of the vehicle and the angle of the blur

kernel in each subregion can be determined by the

vector from the vanishing point to the central point.

The length of the kernel is determined by the vehi-

cle speed and the shutter speed, and we can read from

the camera and the vehicle CAN. In Fig. 8, we can

see that using this simple kernel the disparity result

is similar for the optical ﬂow and cepstrum methods.

We also compare the computation time of these three

algorithms including the image restoration and stereo

vision phase in Table 1 and size of the input stereo

image is 640 × 480.

5 NON-BLIND DEBLURRING

ALGORITHM

After we get the PSF estimation, we deﬁne the fol-

lowing cost function to restore the clear image:

E(l) = ||b −

∑

(ω

∗ l) ⊗ f

+ γ|∇l| (12)

where γ denotes the smoothing parameter.

In practice, we divide the input blurry input into

several overlapping parts and in each part we deblur

the image by solving Eq. 13

E(s

) = ||s

− s

⊗ f

+ γ|∇s

| (13)

We utilize the half-quadratic penalty method (Kr-

ishnan and Fergus, 2009) to recover each sub-region.

We introduce two auxiliary variables u

and u

, and

iterated solving following function:











= F

−1



(

F(∂

)∗F(u

)+F(∂

)∗F(u

)

+F( f )∗F(s

)

(

|F(∂

+|F(∂

)

+|F( f )|



= argmin

γ|u

| + (u

− ∂

)

= argmin

γ|u

| + (u

− ∂

)

(14)

VISAPP 2017 - International Conference on Computer Vision Theory and Applications

406

Figure 8: The ﬁrst column is the original image and the disparity result of MPV (Long et al., 2014b). The second to last

column are the PSF and disparity results of optical ﬂow, cepstrum and the proposed kernel methods, respectively.

(a) (b)

Figure 9: (a) is the blurry and noisy left image. (b) is the

stereo result without noise reduction. (c) is the stereo result

using the denoise algorithm as pre-processing method. (d)

is the stereo result of the proposed method.

After all the sub-regions are restored, we obtain

a weighted sum of these sub-regions with Hanning

window function to produce the ﬁnal results. Since

in the proposed method the blur kernel is not totally

correct in some area, the γ for the proposed method

is 10 times bigger than the optical ﬂow and cepstrum

methods.

5.1 Noisy Case

In the low light condition such as a tunnel, the cap-

tured images are not only blurry but also contain high-

level noise. Simply using the denoising algorithm

as a pre-processing step will improve the quality the

deblurred result, but signiﬁcant artifacts can be ob-

servable along the deblurred image edges and/or the

image structure may be over-smoothed (Tai and Lin,

2012).

Since in our simple case our blur kernel is not cor-

rect in some parts, we already use a large smoothing

parameter to compensate the result. Using the denois-

ing algorithm will further make the ﬁnal result even

more smooth. In this paper, we adopt the follow-

ing strategy to handle this problem as shown in Alg.

1. First, we denoise the input image using fast non-

local means algorithm (Goossens et al., 2010). With

the denoising output

b we compute a deblurring re-

sult

l. Then, we reﬁne

b according to the motion blur

constraints from

l avoid over-sharpening and/or over-

smoothing in the denoised image. Last, we use pre-

vious non-blind algorithm to deblur the reﬁned

b and

get the deblurred image. Speciﬁcally, we compute the

l by minimizing following cost function:

l) = ||b −

∑

(ω

∗

l) ⊗ f

+ λ

||∇l||

(15)

After we get

l, we reﬁne the blurry image

b as:

b) = ||

b −

∑

(ω

∗

l) ⊗ f

+ λ

||∇

b||

+ λ

||b −

b||

(16)

Since Equation 15 and 16 are both l2 norm func-

tion, we can directly compute the closed-form solu-

tion using FFT. In Fig. 9, we show the stereo results

of this noise reduction method. In Fig. 9 (b), we can

see that the stereo result without dealing the noise still

contains noise. In Fig. 9 (c), the non-local means al-

gorithm is adopted to be a pre-processing step and the

stereo result is over-smoothed. In Fig. 9 (d), our re-

sult is relatively better than others with less noises and

correct disparity results for blurry and noisy areas.

6 EXPERIMENT

We evaluated the proposed method on the experimen-

tal autonomous car with the stereo camera mounted

Real-time Stereo Vision System at Tunnel

407

Algorithm 1: Framework for denoising and deblurring.

1: Input: The noisy and blurry image b

2: Output: The denoised and deblurred image l

3: Compute denoised image

b by (Goossens et al.,

2010)

4: Solve

l by minimize Equation 15.

5: Reﬁne

b by minimize Equation 16.

6: Solve l by minimize Equation 12.

Table 2: Result on the tunnel scene.

Stereo Out-Noc(3px) Avg-Noc

Proposed 12.21% 2.12px

SGM (Hirschm

uller, 2008) 32.97% 4.77px

ELAS (Geiger et al., 2011) 20.10% 3.52px

MPV (Long et al., 2014b) 24.33% 4.19px

outside the car and the stereo camera is Bumblebee

BBX3-13S2C-38. We construct a tunnel stereo vision

dataset that contain 200 examples using the Velodyne

HDL-32E 3D laser scanner (10 Hz, 32 laser beams,

range: 100 m) to get precise distance and construct

the ground truth data. We compare the quantitative

values of the proposed algorithm and other real-time

methods in Table 2 and the proposed algorithm out-

performs other existing methods for the tunnel im-

ages. We show some of the qualitative results in Fig.

After computing the disparity, we validate the pro-

posed algorithm within a ADAS application. We

compute the road surface by using the disparity re-

sults. The road surface is detected by using the his-

togram of disparity in horizontal and vertical direc-

tion. We denote the V-disparity (H

) to be the row-

based histogram of the different disparities in each

row of the disparity map and the U-disparity (H

) to

be the column-based histogram. Then we use the V-

disparity and U-disparity (Long et al., 2014a) Viterbi

based optimization method to detect the road surface.

The results are shown in Fig. 10. We can see the dis-

parity results of SGM (Hirschm

uller, 2008) contain

much more holes and errors in the road surface, and

generate a slightly more irregular road area and road

boundary.

6.1 Extend to Non-tunnel Case

We evaluate the proposed algorithm on some non-

tunnel images in low light conditions such as night-

time, the captured images will also be blurred and

noisy. Although the background of the nighttime im-

ages are more complicated than the tunnel case and

the blur kernel of the proposed method is not suit-

able in some parts of the image, the proposed algo-

rithm still can signiﬁcantly improve the stereo results

Figure 10: Road surface comparison. The ﬁrst to last

columns represent the disparity result and corresponding

road surface estimation of method (Long et al., 2014b),

(Hirschm

uller, 2008) and the propose simple method.

Figure 11: Results on the nighttime images. The ﬁrst to last

columns represent is the input left image of stereo pair, the

stereo results of (Long et al., 2014b) and our results.

as shown in Fig. 11.

7 CONCLUSION

In this paper, we have presented a novel algorithm to

improve stereo vision in the tunnel. Typically, the mo-

tion kernel for tunnel cases is spatially-variant and

difﬁcult to be estimated in the real-time. Consid-

ering the characteristics of the tunnel case, we seg-

ment the image into overlapped subregions and as-

sume the motion kernel in each region to be a lin-

ear function. Then we utilize three algorithms to es-

timate the PSF by optical ﬂow, cepstrum and simple

constant kernel. We compare the results of these three

approaches and ﬁnd that the proposed simple kernel is

enough to produce promising disparity results with re-

duced computational efﬁciency. This simple method

together with the fast non-blind deblurring algorithm

and multi-scale MPV method are the ﬁnal real-time

deblurring-stereo framework for the tunnel case. The

experiment results show the proposed algorithm can

effectively and efﬁciently deal with the real tunnel im-

ages. Furthermore although the proposed algorithm is

designed for tunnel images, it performs well on other

images captured at low light conditions.

VISAPP 2017 - International Conference on Computer Vision Theory and Applications

408

REFERENCES

Boykov, Y., Veksler, O., and Zabih, R. (2001). Fast ap-

proximate energy minimization via graph cuts. Pat-

tern Analysis and Machine Intelligence, IEEE Trans-

actions on, 23(11):1222–1239.

Cho, S. and Lee, S. (2009). Fast motion deblurring. In

ACM Transactions on Graphics (TOG), volume 28,

page 145. ACM.

Fergus, R., Singh, B., Hertzmann, A., Roweis, S. T., and

Freeman, W. T. (2006). Removing camera shake from

a single photograph. In ACM Transactions on Graph-

ics (TOG), volume 25, pages 787–794. ACM.

Geiger, A., Roser, M., and Urtasun, R. (2011). Efﬁcient

large-scale stereo matching. In Computer Vision–

ACCV 2010, pages 25–38. Springer.

Goossens, B., Luong, H., Aelterman, J., Pi

zurica, A.,

and Philips, W. (2010). A gpu-accelerated real-

time nlmeans algorithm for denoising color video se-

quences. In International Conference on Advanced

Concepts for Intelligent Vision Systems, pages 46–57.

Springer.

He, K., Sun, J., and Tang, X. (2013). Guided image ﬁlter-

ing. Pattern Analysis and Machine Intelligence, IEEE

Transactions on, 35(6):1397–1409.

Hirsch, M., Schuler, C. J., Harmeling, S., and Scholkopf, B.

(2011). Fast removal of non-uniform camera shake.

In Computer Vision (ICCV), 2011 IEEE International

Conference on, pages 463–470. IEEE.

Hirschm

uller, H. (2008). Stereo processing by semiglobal

matching and mutual information. Pattern Analy-

sis and Machine Intelligence, IEEE Transactions on,

30(2):328–341.

Joshi, N., Kang, S. B., Zitnick, C. L., and Szeliski, R.

(2010). Image deblurring using inertial measure-

ment sensors. ACM Transactions on Graphics (TOG),

29(4):30.

Krishnan, D. and Fergus, R. (2009). Fast image deconvo-

lution using hyper-laplacian priors. In Advances in

Neural Information Processing Systems, pages 1033–

1041.

Long, Q., Xie, Q., Mita, S., Ishimaru, K., and Shirai, N.

(2014a). A real-time dense stereo matching method

for critical environment sensing in autonomous driv-

ing. In Intelligent Transportation Systems (ITSC),

2014 IEEE 17th International Conference on, pages

853–860. IEEE.

Long, Q., Xie, Q., Mita, S., Tehrani, H., Ishimaru, K., and

Guo, C. (2014b). Real-time dense disparity estima-

tion based on multi-path viterbi for intelligent vehicle

applications. In Proceedings of the British Machine

Vision Conference. BMVA Press.

Rom, R. (1975). On the cepstrum of two-dimensional func-

tions (corresp.). Information Theory, IEEE Transac-

tions on, 21(2):214–217.

Scharstein, D. and Szeliski, R. (2002). A taxonomy and

evaluation of dense two-frame stereo correspondence

algorithms. International journal of computer vision,

47(1-3):7–42.

Son, T. T. and Seiichi, M. (2006). Stereo matching algo-

rithm using a simpliﬁed trellis diagram iteratively and

bi-directionally. IEICE transactions on information

and systems, 89(1):314–325.

Sun, J., Zheng, N.-N., and Shum, H.-Y. (2003). Stereo

matching using belief propagation. Pattern Analy-

sis and Machine Intelligence, IEEE Transactions on,

25(7):787–800.

Tai, Y.-W. and Lin, S. (2012). Motion-aware noise ﬁltering

for deblurring of noisy and blurry images. In Com-

puter Vision and Pattern Recognition (CVPR), 2012

IEEE Conference on, pages 17–24. IEEE.

Tai, Y.-W., Tan, P., and Brown, M. S. (2011). Richardson-

lucy deblurring for scenes under a projective mo-

tion path. Pattern Analysis and Machine Intelligence,

IEEE Transactions on, 33(8):1603–1618.

Veksler, O. (2005). Stereo correspondence by dynamic pro-

gramming on a tree. In Computer Vision and Pattern

Recognition, 2005. CVPR 2005. IEEE Computer So-

ciety Conference on, volume 2, pages 384–390. IEEE.

Wang, R. and Tao, D. (2014). Recent progress in image

deblurring. arXiv preprint arXiv:1409.6838.

Whyte, O., Sivic, J., and Zisserman, A. (2011). Deblurring

shaken and partially saturated images. In Computer

Vision Workshops (ICCV Workshops), 2011 IEEE In-

ternational Conference on, pages 745–752. IEEE.

Xu, L. and Jia, J. (2010). Two-phase kernel estimation for

robust motion deblurring. In Computer Vision–ECCV

2010, pages 157–170. Springer.

Xu, Y., Hu, X., Wang, L., and Peng, S. (2012). Single im-

age blind deblurring with image decomposition. In

Acoustics, Speech and Signal Processing (ICASSP),

2012 IEEE International Conference on, pages 929–

932. IEEE.

Zabih, R. and Woodﬁll, J. (1994). Non-parametric local

transforms for computing visual correspondence. In

Computer VisionłECCV’94, pages 151–158. Springer.

Zbontar, J. and LeCun, Y. (2014). Computing the stereo

matching cost with a convolutional neural network.

arXiv preprint arXiv:1409.4326.

Zheng, S., Xu, L., and Jia, J. (2013). Forward motion de-

blurring. In Proceedings of the IEEE International

Conference on Computer Vision, pages 1465–1472.

Real-time Stereo Vision System at Tunnel

409