Superpixels in Pedestrian Detection from Stereo Images in

Urban Traffic Scenarios

Ion Giosan and Sergiu Nedevschi

Computer Science Department, Technical University of Cluj-Napoca, Cluj-Napoca, Romania

Keywords: Superpixels, Pedestrian Hypotheses, HOG Features, PCA, SVM, AdaBoost, Hypotheses Validation,

Speed-up Detection.

Abstract: Pedestrian detection is a common task in every driving assistance system. The main goal resides in obtaining

a high accuracy detection in a reasonable amount of processing time. This paper proposes a novel method for

superpixel-based pedestrian hypotheses generation and their validation through feature classification. We

analyze the possibility of using superpixels in pedestrian detection by investigating both the execution time

and the accuracy of the results. Urban traffic images are acquired by a stereo-cameras system. A multi-feature

superpixels-based method is used for obstacles segmentation and pedestrian hypotheses selection. Histogram

of Oriented Gradients features are extracted both on the raw 2D intensity image and also on the superpixels

mean intensity image for each hypothesis. Principal Component Analysis is also employed for selecting the

relevant features. Support Vector Machine and AdaBoost classifiers are trained on: initial features and

selected features extracted from both raw 2D intensity image and mean superpixels intensity image. The

comparative results show that superpixels- based pedestrian detection clearly reduce the execution time while

the quality of the results is just slightly decreased.

1 INTRODUCTION

The high rate of road accidents in the world has

motivated the development of more efficient driving

assistance systems in order to reduce the number of

accidents with injuries and fatalities. The driving

assistance systems that have implemented functions

like obstacle detection and obstacle recognition

should be robust in various traffic and weather

conditions. Although highway traffic scenes are

relatively simple mainly containing obstacles such as

cars, fences and poles, urban traffic scenarios are

much more complex. In urban traffic scenarios

several problems occur due to the crowded scenes in

which occlusions occur frequently and makes the

obstacle detection problem very difficult. Here the

pedestrians, which are the most vulnerable traffic

participants, appear.

Although pedestrian recognition is a simple

problem for humans, it is a relatively difficult

problem for artificial vision systems. This is mainly

due to the different possible traffic situations,

different clothing and accessories they wear. Usually,

the pedestrians are recognized in traffic scenes by

using obstacles classifiers trained on specific features.

Stereo-cameras are frequently used in computer

vision at the expense of other sensors. We use a

solution consisting of a gray level stereo-cameras

setup that offers the possibility of exploiting both the

2D intensity and the associated 3D points’

information.

A very important and motivating aspect for

driving assistance systems is that all the processing

must be done as quickly as possible in order to

achieve real-time execution. This is particularly

important because it must give a timely response at

high speeds of ego-vehicle. The obstacle detection

and recognition algorithms must be complex in order

to robustly solve problems but they must be very fast

for delivering timely results. These two requirements,

that are in a continuous compromise, must be met

both as much as possible.

We proposed a novel generation and validation of

pedestrian hypotheses based on superpixels. The goal

resides in reducing the processing time for pedestrian

detection while preserving the accuracy of the results.

We use our previous work in superpixels-based

obstacle segmentation (Giosan and Nedevschi, 2014)

in order to accurately determine the traffic pedestrian

hypotheses. Classic features like Histogram of

Giosan, I. and Nedevschi, S.

Superpixels in Pedestrian Detection from Stereo Images in Urban Trafﬁc Scenarios.

DOI: 10.5220/0005680305010508

In Proceedings of the 11th Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2016) - Volume 4: VISAPP, pages 501-508

ISBN: 978-989-758-175-5

501

Oriented Gradients (HOG) are computed both on the

raw 2D intensity image and also on the superpixels

mean image for comparison purposes. Principal

Component Analysis (PCA) is also employed for

feature dimensionality reduction. Traditional

classifiers like Support Vector Machine (SVM) and

AdaBoost are trained and tested in order to prove that

using superpixels for hypotheses validation can

strongly reduce the processing time while the quality

of the pedestrian detection results is not so affected.

The rest of the paper is organized as follows: in

section 2 we present the related work, in section 3 the

system overview, in section 4 the superpixel-based

pedestrian hypotheses detection, in section 5 the

extracted features and the classifiers that are used, in

section 6 the experimental results and finally in

section 7 we draw the conclusions of this work.

2 RELATED WORK

Superpixels are clusters of pixels based on local

image features. SLIC superpixels described in

(Achanta et al., 2012) represent a fast approach that

can be used for segmenting gray levels images in

separate superpixels. They may be used for reducing

the complexity of subsequent image processing tasks

like obstacle detection. Pedestrian hypotheses are

usually extracted from the set of detected obstacles by

imposing some pedestrian specific geometrical

constraints. The hypotheses are then used for

reducing the search space, resulting in a faster

pedestrian detection process.

Usually stereovision based approaches are widely

used first in traffic scenes obstacle segmentation

(Oniga and Nedevschi, 2010) and second for

validating the obstacle classification results (Bertozzi

et al., 2008). Features extraction and feature-based

classifiers represent intermediate steps in obstacles

classification. Features are usually extracted from the

2D appearance obstacle images but they can also

integrate depth information and optical flow motion

information. Obstacle classifiers may be trained

directly on the extracted features or on a subset of

relevant features (You and Ruichek, 2012). A high

quality of the stereo-reconstruction process (Pantilie

and Nedevschi, 2012) is absolutely necessary for

obtaining a dense and accurate 3D points map. Based

on this map, several algorithms like points grouping

(Pocol et al., 2008) or density map analysis

(Nedevschi et al., 2009) may be used for obstacle

segmentation. In comparison with monocular vision

based techniques that uses symmetry (Bertozzi et al.,

2000), edges (Bertozzi and Broggi, 1998) and

textures (Heikkila and Pietikainen, 2006) from

intensity information, stereovision based obstacle

segmentation approaches (Broggi et al., 2011,

Nedevschi et al., 2004, Llorca et al., 2012) are clearly

superior.

Methods that divide the image pixels into regions

having the properties that all pixels from a separate

region are similar with respect to a chosen similarity

metric are presented in (Felzenszwalb and

Huttenlocher, 2004) and (Xiaofeng and Malik, 2003).

A graph where the nodes are the image pixels and the

edges represent a neighborhood relationship between

pixels is computed. These methods represent the basis

of the, nowadays very common and superior,

superpixels based image segmentation approaches. In

(Giosan and Nedevschi, 2014) we proposed a novel

obstacle detection method based on the original scene

segmentation in superpixels. The method combined

the intensity, depth and motion information within the

SLIC superpixels. A novel algorithm was proposed

for superpixels clustering into obstacles and obstacles

refinement. A method for very close obstacles

separation was developed based on the motion

vectors analysis of their component superpixels. The

results showed a very good obstacles detection with

precise segmentation of their surfaces which is

particularly useful for subsequent processes like

pedestrian detection. Continuing this work, in this

paper we propose a novel method for superpixels-

based generation and validation of pedestrian

hypotheses. The superpixels benefits in the pedestrian

detection process are clearly highlighted.

In the literature, several methods use different

discriminant features like shapes and edges (Broggi

et al., 2000), contours (Hilario et al., 2005), contour

templates (Gavrila and Philomin, 1999; Gavrila,

2000; Giosan and Nedevschi, 2009), symmetries

(Havasi et al., 2004), Haar features (Papageorgiou

and Poggio, 2000), HOG features (Dalal and Triggs,

2005) used for pedestrian detection. Usually these

features are firstly extracted on pedestrian hypotheses

and then fed into classifiers that are able to distinguish

between pedestrians and other traffic scene obstacles.

A lot of different methods exist for feature based

obstacle and specifically pedestrian classification

(Giosan and Nedevschi, 2012). In (Rivlin et al., 2002,

Lun et al., 2007), a SVM classifier is used for

recognizing pedestrians and bikes in traffic scenes. A

powerful Adaboost classifier built upon some

characteristics of rectangular edge description is

proposed in (Yi et al., 2010) for high accuracy

pedestrian recognition. Neural networks are also used

for high-accuracy pedestrian and other obstacles

classification (Toth and Aach, 2003). The state of the

VISAPP 2016 - International Conference on Computer Vision Theory and Applications

502

art in pedestrian detection achieve high accuracy

results even from monocular images (Benenson et al.,

2012). The benefits of stereo images in regions of

interest generation and localization and an evaluation

methodology of pedestrian detection is described in

(Keller et al., 2011).

3 SUPERPIXELS BASED

PEDESTRIAN DETECTION

SYSTEM OVERVIEW

Figure 1: Superpixels-based pedestrian detection system

overview.

The structural architecture of the entire system

and the flow between its modules that is used for

analyzing the influence of superpixels in pedestrian

detection from urban traffic scenes is presented in

Figure 1.

Traffic scenes images are acquired with a

stereovision cameras setup. These consist of

grayscale image sequences with resolution of

512x383 pixels. The stereo-reconstruction process is

achieved by the SORT-SGM algorithm (Pantilie and

Nedevschi, 2012) which offers an accurate and dense

depth map, crucial for further processing steps.

Optical flow is computed by using the Lukas-Kanade

approach on good features to track (Shi and Tomasi,

1994). Combining intensity, depth and optical flow

information, our superpixels based obstacle detector

(Giosan and Nedevschi, 2014) is used for selecting

the pedestrian hypotheses. HOG features are

extracted both directly on raw 2D and on the

superpixels hypotheses intensity ROI images. PCA

procedure is or not applied before training two robust

classifiers: SVM and AdaBoost. These classifiers are

trained on a database containing pedestrians and non-

pedestrians instances. Finally, the classification

results from all investigated ways are analyzed and

the benefits of using superpixels for pedestrian

detection are highlighted.

4 SUPERPIXELS-BASED

PEDESTRIAN HYPOTHESES

DETECTION

We take advantage of our previous work (Giosan and

Nedevschi, 2014) for superpixels based obstacle

detection. SLIC superpixels are computed in a

rectangular region of interest (ROI) defined by the

positions and dimensions: (left=0, top=100,

right=512, bottom=320) (see Figura). A fixed number

of N=2000 superpixels segments the established

intensity image ROI. A set of intensity, depth and

motion features are used in order to cluster the

superpixels into scene obstacles (see Figurb).

A modified version of the novel obstacle detection

algorithm described by us in (Giosan and Nedevschi,

2014) is proposed for finding only the pedestrian

hypotheses that will be further used in the validation

(pedestrian detection) process. The algorithm has five

main steps, the last one addressing directly the

pedestrian hypotheses selection from the traffic scene

segmented obstacles:

 Gray-levels SLIC superpixels computation

 Superpixels specific features extraction

 Superpixels clustering in obstacles hypotheses

 Obstacles hypotheses validation and refinement

 Pedestrian hypotheses generation

Superpixels in Pedestrian Detection from Stereo Images in Urban Trafﬁc Scenarios

503

Figure 2: Obstacles detection: a) intensity image ROI; b)

corresponding superpixels-based detected obstacles (with

random color for each obstacle).

Pedestrian hypotheses are considered to be those

superpixels-based obstacles that meet some

geometrical constraints. We define the following

constraints for the surrounding 3D cuboid that are

satisfied for the majority of pedestrians:

 Height between 1.5m and 2.2m

 Width between 0.3m and 0.9m

 Length between 0.3m and 0.9m

 2D image ROI aspect ratio (height/width)

between 1.0 and 4.0

Figure 3: Superpixels-based pedestrian hypotheses (with

random color for each obstacle).

In case of traffic scene image presented in Figure

2, the generated superpixel-based pedestrian

hypotheses are shown in Figure 3.

5 HOG FEATURES

EXTRACTION, SELECTION

AND CLASSIFIERS TRAINING

We choose the HOG features for analyzing the

influence of superpixels in the pedestrian detection

(hypotheses validation) process. These features are

extracted both on the raw 2D intensity ROIs and on

the superpixels ROIs of the pedestrian hypotheses. A

PCA feature dimensionality reduction method is also

investigated in speeding up the final classification

while preserving the accuracy. Different classifiers

are trained both on the entire HOG features set and on

PCA selected features. The features are extracted on

a training set with pedestrians and non-pedestrians

intensity images.

The raw 2D intensity image ROIs (see Figure 4a)

of the pedestrian hypotheses are firstly resized to a

fixed dimension of 64x128 pixels using a fast bilinear

interpolation procedure (see Figure 4b). The resized

image is divided in blocks with 16x16 pixels and each

block in 8x8 pixels cells. We choose 9 bins for the

HOG computation. The resulting HOG feature vector

for each instance has a number of 3780 components.

a) b) c) d)

Figure 4: Pedestrian hypotheses: a) raw 2D intensity ROI;

b) resized 2D intensity ROI; c) superpixels on resized 2D

intensity ROI; d) mean superpixels image.

We choose a SVM and an AdaBoost classifier

trained using several feature vectors extracted for

pedestrians and non-pedestrians. We also reduce the

number of features by keeping only the first 30

principal components out of all 3780 initial features

and retrain the previously mentioned classifiers.

In order to investigate the possible advantages of

using superpixels in pedestrian detection, we reduce

VISAPP 2016 - International Conference on Computer Vision Theory and Applications

504

the raw 2D intensity ROIs by resizing it to 16x32

pixels and extract the SLIC superpixels (see Figure

4c). We compute the mean superpixels image (see

Figure 4d) by averaging all the intensities at every

superpixel level. HOG features are extracted, using

the same block size, cell size and number of bins, on

the superpixels mean image resulting in a 108 feature

vector. Identically to the training process on the raw

features previously described, SVM and AdaBoost

classifiers are trained first on several instances with

all 108 feature vectors and then by selecting only the

first 30 principal components.

The classification performance and the processing

time obtained by the SVM and AdaBoost classifiers

trained on extracted HOG features on initial image

and on superpixels mean image, with and without

PCA are presented comparatively in the Experimental

Results section.

6 EXPERIMENTAL RESULTS

In this section we describe the obtained results for

both generation and validation of pedestrian

hypotheses. We also highlight the advantages and

disadvantages of using superpixels.

Pedestrian hypotheses generation was tested on

image sequences from common urban traffic

scenarios, where the results seems to be similar to

those obtained by other obstacle detection methods.

In difficult traffic scenarios with crowded and

hanging obstacles, the superpixels-based pedestrian

hypotheses generation is clearly superior. A

comparison of our method with the pedestrian

hypotheses generation approach based on 3D points

grouping (Pocol et al., 2008) and density maps

(Nedevschi et al., 2009) is presented in Figure. The

superpixels-based proposed method offers better

hypotheses generation, reducing the number of false

detections. It also provides better shape definition

which is very important for further validation in the

pedestrian detection process.

In the evaluation of the pedestrian hypotheses

validation process, we use obstacles instances from

our own database containing about 25,000 intensity

images of pedestrians and another 25,000 intensity

images of non-pedestrians (see Figure).

We trained two classifiers: SVM with RBF

kernels and AdaBoost with decision stumps weak

learners. All the training procedures were conducted

by using a number of maximum 5000 instances

selected randomly from our database. The same

number of other instances is also selected randomly

and used for evaluating the classification accuracy.

The processing time is also counted for each

experiment.

Figure 5: Pedestrian hypotheses generation: a) by means of

3D points grouping and density maps (pedestrians from the

left and right side are wrongly grouped together); b) using

the proposed superpixels-based method (pedestrians from

the left and right side are clearly separated).

Figure 6: Sample dataset images: pedestrians and non-

pedestrians.

The SVM classifier performance obtained in all

four proposed scenarios

 I: no superpixels, no PCA

 II: with superpixels, no PCA

 III: no superpixels, with PCA

 IV: with superpixels, with PCA

are briefly presented in Table 1.

Table 1: SVM Classifier’s Performance.

Scenario

Number of

test instances

False Positive

Rate

True

Positive

Rate

Processing

time (ms)

I 5000 0.044 0.937 58298

II 5000 0.078 0.898 6499

III 5000 0.056 0.91 5310

IV 5000 0.081 0.868 5605

Superpixels in Pedestrian Detection from Stereo Images in Urban Trafﬁc Scenarios

505

Plotting the false positive rates and true positive

rates for pedestrian detection in all four proposed

training/testing scenarios show that there is small

decrease in true positive rates of about 4% (see Figure

7) with a small increase in false positive rates of about

3% (see Figure 8

Figure) when considering superpixels

mean image versus raw image processing.

Figure 7: SVM classifier pedestrian detection true positive

rates in all four scenarios: with/without superpixels (SP),

with/without PCA.

Figure 8: SVM classifier pedestrian detection false positive

rates in all four scenarios: with/without superpixels (SP),

with/without PCA.

The main advantage is depicted in Figure where

the processing time interval seems to be very short

(decreased with about 90%) when processing PCA

superpixels images versus raw intensity images in

pedestrian detection.

The same procedures like in the SVM classifier are

applied for training and testing the AdaBoost

classifier. The AdaBoost classifier performance

obtained in the same four proposed scenarios are

briefly presented in Table 2.

Table 2: AdaBoost Classifier’s Performance.

Scenario

Number of

test instances

False Positive

Rate

True

Positive

Rate

Processing

time (ms)

I 5000 0.1136 0.893 7544

II 5000 0.147 0.826 5090

III 5000 0.128 0.872 8660

IV 5000 0.134 0.806 4055

Figure 9: AdaBoost classifier pedestrian detection true

positive rates in all four scenarios: with/without superpixels

(SP), with/without PCA.

Figure 10: AdaBoost classifier pedestrian detection false

positive rates in all four scenarios: with/without superpixels

(SP), with/without PCA.

As in the case of SVM classifier, using the

AdaBoost classifier in all four proposed

training/testing scenarios show that there is small

decrease in true positive rates of about 6% (see Figure

9) with a small increase in false positive rates of about

3% (see Figure 10) when considering superpixels

mean image versus raw image processing. The main

advantage is depicted in Figure 11 where the

processing time seems to be shorter (with about 40%)

when processing PCA superpixels images versus raw

intensity images for pedestrian detection.

The experimental results were obtained by

processing images containing pedestrians and non-

pedestrians acquired from different traffic scenarios.

The initial superpixels-based obstacle segmentation

process was performed by our own novel proposed

algorithms previously published in (Giosan and

Nedevschi, 2014). Generation and validation of

pedestrian hypotheses were achieved by the novel

described methods in this paper. The system achieves

real-time performance running on a single core of an

Intel Core i7-4790 processor @ 3.60 GHz.

VISAPP 2016 - International Conference on Computer Vision Theory and Applications

506

Figure 11: SVM and AdaBoost classifiers processing time:

comparison between the extreme scenarios I and IV.

7 CONCLUSIONS

A novel approach for superpixel-based generation

and validation of pedestrian hypotheses was

described. The proposed superpixels-based

pedestrian hypotheses generation method offers

better hypotheses generation results even in difficult

traffic scenarios. It clearly reduces the number of

false hypotheses detections and provides better

pedestrians shape definition. The influence of using

superpixels in pedestrian detection (hypotheses

validation) was also investigated. Classical HOG

features were computed both on the raw 2D intensity

image and also on the superpixels mean image. PCA

feature dimensionality reduction was also employed

for speeding up the classification process while trying

to preserve the accuracy. SVM and AdaBoost

classifiers were trained and tested on random

pedestrian/non-pedestrian intensity images. Their

classification results were compared in terms of both

accuracy and also in terms of processing time

interval.

The SVM classification proved that the

processing time interval may be decreased with 90%

when using superpixels, while the accuracy lost only

3% in false positive rate. The AdaBoost classification

proved that the processing time may be decreased

with 40% when using superpixels while the accuracy

lost also only 3% in false positive rate. The

superpixels usage clearly improves the quality of the

pedestrian hypotheses generation. Hypotheses

validation (pedestrian detection) shows a significant

decrease of the processing time while the small loss

in accuracy can be neglected.

REFERENCES

Achanta, R., Shaji, A., Smith, K., Lucchi, A., Fua, P. &

Susstrunk, S. 2012. SLIC Superpixels Compared to

State-of-the-Art Superpixel Methods. IEEE

Transactions on Pattern Analysis and Machine

Intelligence, 34, pp. 2274-2282.

Benenson, R., Mathias, M., Timofte, R. & Van Gool, L.

2012. Pedestrian detection at 100 frames per second.

IEEE Conference on Computer Vision and Pattern

Recognition (CVPR). pp. 2903-2910.

Bertozzi, M., Bombini, L., Cerri, P., Medici, P., Antonello,

P. C. & Miglietta, M. 2008. Obstacle detection and

classification fusing radar and vision. IEEE Intelligent

Vehicles Symposium. pp. 608-613.

Bertozzi, M. & Broggi, A. 1998. GOLD: a parallel real-time

stereo vision system for generic obstacle and lane

detection. IEEE Transactions on Image Processing, 7,

pp. 62-81.

Bertozzi, M., Broggi, A., Fascioli, A. & Nichele, S. 2000.

Stereo vision-based vehicle detection. Proceedings of

the IEEE Intelligent Vehicles Symposium. pp. 39-44.

Broggi, A., Bertozzi, M., Fascioli, A. & Sechi, M. 2000.

Shape-based pedestrian detection. IEEE Intelligent

Vehicles Symposium. pp. 215-220.

Broggi, A., Buzzoni, M., Felisa, M. & Zani, P. 2011. Stereo

obstacle detection in challenging environments: The

VIAC experience. International Conference on

Intelligent Robots and Systems (IROS). pp. 1599-1604.

Dalal, N. & Triggs, B. 2005. Histograms of oriented

gradients for human detection. IEEE Computer Society

Conference on Computer Vision and Pattern

Recognition. pp. 886-893 vol. 1.

Felzenszwalb, P. F. & Huttenlocher, D. P. 2004. Efficient

Graph-Based Image Segmentation. International

Journal of Computer Vision, 59, pp. 167-181.

Gavrila, D. 2000. Pedestrian Detection from a Moving

Vehicle. Proceedings of the 6th European Conference

on Computer Vision-Part II, pp. 37-49.

Gavrila, D. M. & Philomin, V. 1999. Real-time object

detection for "smart" vehicles. Proceedings of the

Seventh IEEE International Conference on Computer

Vision. pp. 87-93 vol.1.

Giosan, I. & Nedevschi, S. 2009. Building Pedestrian

Contour Hierarchies for Improving Detection in Traffic

Scenes. Computer Vision and Graphics - Lecture Notes

in Computer Science, pp. 154-163.

Giosan, I. & Nedevschi, S. 2012. A solution for

probabilistic inference and tracking of obstacles

classification in urban traffic scenarios. IEEE

International Conference on Intelligent Computer

Communication and Processing (ICCP). pp. 221-227.

Giosan, I. & Nedevschi, S. 2014. Superpixel-based obstacle

segmentation from dense stereo urban traffic scenarios

using intensity, depth and optical flow information.

IEEE 17th International Conference on Intelligent

Transportation Systems (ITSC). pp. 1662-1668.

Havasi, L., Szlavik, Z. & Sziranyi, T. 2004. Pedestrian

Detection Using Derived Third-Order Symmetry of

Legs. Proceedings of the IEEE International

Superpixels in Pedestrian Detection from Stereo Images in Urban Trafﬁc Scenarios

507

Conference on Computer Vision and Graphics, pp. 11-

17.

Heikkila, M. & Pietikainen, M. 2006. A Texture-Based

Method for Modeling the Background and Detecting

Moving Objects. IEEE Transactions on Pattern

Analysis and Machine Intelligence, 28, pp. 657-662.

Hilario, C., Collado, J. M., Armingol, J. M. & Escalera, A.

D. L. 2005. Pedestrian detection for intelligent vehicles

based on active contour models and stereo vision.

Proceedings of the 10th international conference on

Computer Aided Systems Theory, pp. 537-542.

Keller, C. G., Enzweiler, M. & Gavrila, D. M. 2011. A new

benchmark for stereo-based pedestrian detection. IEEE

Intelligent Vehicles Symposium (IV). pp. 691-696.

Llorca, D. F., Sotelo, M. A., Hellín, A. M., Orellana, A.,

Gavilan, M., Daza, I. G. & Lorente, A. G. 2012. Stereo

regions-of-interest selection for pedestrian protection:

A survey. Transportation research part C: emerging

technologies, 25, pp. 226-237.

Lun, Z., Li, S. Z., Xiaotong, Y. & Shiming, X., 2007. Real-

time Object Classification in Video Surveillance Based

on Appearance Learning. IEEE Conference on

Computer Vision and Pattern Recognition. pp. 1-8.

Nedevschi, S., Bota, S. & Tomiuc, C., 2009. Stereo-Based

Pedestrian Detection for Collision-Avoidance

Applications. IEEE Transactions on Intelligent

Transportation Systems, 10, pp. 380-391.

Nedevschi, S., Danescu, R., Frentiu, D., Marita, T., Oniga,

F., Pocol, C., Schmidt, R. & Graf, T. 2004. High

accuracy stereo vision system for far distance obstacle

detection. IEEE Intelligent Vehicles Symposium. pp.

292-297.

Oniga, F. & Nedevschi, S. 2010. Processing Dense Stereo

Data Using Elevation Maps: Road Surface, Traffic Isle,

and Obstacle Detection. IEEE Transactions on

Vehicular Technology, 59, pp. 1172-1182.

Pantilie, C. D. & Nedevschi, S. 2012. SORT-SGM:

Subpixel Optimized Real-Time Semiglobal Matching

for Intelligent Vehicles. IEEE Transactions on

Vehicular Technology, 61, pp. 1032-1042.

Papageorgiou, C. & Poggio, T. 2000. A Trainable System

for Object Detection. International Journal of

Computer Vision, 38, pp. 15-33.

Pocol, C., Nedevschi, S. & Meinecke, M. M. 2008.

Obstacle Detection Based on Dense Stereovision for

Urban ACC Systems. International Workshop on

Intelligent Transportation (WIT), Hamburg, Germany.

pp. 13-18.

Rivlin, E., Rudzsky, M., Goldenberg, R., Bogomolov, U. &

Lepchev, S. 2002. A real-time system for classification

of moving objects. 16th International Conference on

Pattern Recognition. pp. 688-691 vol.3.

Shi, J. & Tomasi, C. 1994. Good features to track. IEEE

Conference on Computer Vision and Pattern

Recognition. pp. 593-600.

Toth, D. & Aach, T. 2003. Detection and recognition of

moving objects using statistical motion detection and

Fourier descriptors. 12th International Conference on

Image Analysis and Processing. pp. 430-435.

Xiaofeng, R. & Malik, J. 2003. Learning a classification

model for segmentation. Proceedings of IEEE

International Conference on Computer Vision. pp. 10-

17 vol.1.

Yi, T., Wei-Ming, L. & Wu, J. 2010. An improved

pedestrian detection algorithm based on Adaboost

cascading stucture. Proceedings of World Congress on

Intelligent Control and Automation (WCICA). pp.

6321-6326.

You, L. & Ruichek, Y. 2012. Moving objects detection and

recognition using sparse spatial information in urban

environments. IEEE Intelligent Vehicles Symposium.

pp. 1060-1065.

VISAPP 2016 - International Conference on Computer Vision Theory and Applications

508