Pedestrian Detection using HOG-based Block Selection

Minsung Kang and Young Chul Lim

IT Convergence Research Division, DGIST, Daegu, Hyeonpung Myeon, Korea

Keywords: Pedestrian Detection, Intelligent Vehicle, HOG, Camera, Computer Vision.

Abstract: Recently, pedestrian detection methods have been popularly used in the field of intelligent vehicles. In most

previous works, the Histogram of Oriented Gradients (HOG) is used to extract features for pedestrian

detection. However HOG is difficult to use in the real-time operating system of an intelligent vehicle. In this

paper, we proposed a pedestrian detection method using a HOG-based block selection. First, we analyse the

HOG block and select the parts of the block with a high hit rate. We then use only 20% of the total HOG

blocks for the pedestrian feature. The proposed method is 5 times faster than methods using the entire

feature, while performance remains almost the same.

1 INTRODUCTION

Pedestrian detection methods have been used

recently for intelligent vehicle, intelligent robot and

video security applications. Pedestrian detection is

the technical methodology for finding the position of

pedestrians from a camera image. In the detection

process, first, a feature is extracted for pedestrian

classification. Then a pedestrian is detected using a

feature from a searched image. The performance and

computation speed are typically different when

using features that are extracted with various shapes.

The Histogram of Oriented Gradients (HOG) is

one of the well-known features used for pedestrian

detection. The HOG feature is robust to variations of

illumination. However, The HOG feature needs a

high amount of image processing because the

dimensions of the feature are high. Hence,

pedestrian detection based on HOG is impractical

for the real time operation of vehicles. In an

intelligent vehicle, real time operation is important

because reaction time is directly connected to the

safety of the driver and pedestrian.

As a result, many pedestrian detection methods

based on the HOG feature are being researched with

the goal of reducing computation time. Many of

these existing methods change the process of

searching the image to reduce computation time.

Other methods use a GPU to improve computation

speed but these need an NVIDIA graphic card. Such

methods use the high dimensions of HOG and

improve computation speed in post-processing.

However these methods do not solve the

fundamental problem.

Accordingly, this paper proposes a pedestrian

detection method using HOG-based block selection.

The structure of this paper is as follows: Chapter 2

introduces related works about pedestrian detection.

Chapter 3 describes the proposed algorithm. Chapter

4 deals with the verification of the proposed

algorithm through experiments. And finally Chapter

5 presents conclusions.

2 RELATED WORKS

Pedestrian detection methods involve the extraction

of features from a pedestrian dataset and a training

feature using a classifier such as SVM. Then the

feature is used to detect the pedestrian in a whole

camera image. Existing methods typically change

the process of searching the image to reduce

computation time. As shown in Figure 1, the sliding

window method makes an image pyramid from the

original image in order to search the image.

The computation speed of the sliding window

method is very slow because the area being searched

is big. The classifier of the sliding window method

is fixed. Hence, to address this issue, as shown in

Figure 2, an alternative method makes various sizes

of classifier to improve the search speed. But this

method is hard to use because HOG is an invariant

feature. For this reason, a hybrid method has been

proposed, as shown in Figure 3. The computational

783

Kang M. and Lim Y..

Pedestrian Detection using HOG-based Block Selection.

DOI: 10.5220/0005147607830787

In Proceedings of the 11th International Conference on Informatics in Control, Automation and Robotics (IVC&ITS-2014), pages 783-787

ISBN: 978-989-758-040-6

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

Figure 1: Dense image pyramid.

Figure 2: Classifier pyramid.

Figure 3: Hybrid approach.

time is improved but this method results in the loss

of data and its structure is highly complex.

Another method uses a GPU. That method uses a

parallel calculated block histogram for improving

computation time. However, it depends on having

equipment available because it requires an NVIDIA

graphic card. The method uses all of the blocks of

the HOG for pedestrian detection.

The method proposed in this paper, in contrast,

does not use all the blocks of the HOG. As shown in

Figure 4, this new method selects only a part of the

blocks. Figure 4-(a) shows the extracted feature

based on HOG from a general pedestrian database.

The figure of the pedestrian does not fill the entire

Figure 4: Proposed algorithm; (a) HOG feature, (b)

Analysis feature (green: pedestrian, red: another), (c)

Select feature.

Figure 5: HOG (a) original image, (b) oriented gradients,

HOG.

feature. Accordingly, the proposed method performs

an analysis to determine which part of the block

contains a high probability of the pedestrian

existing, as in Figure 4-(b). Then the proposed

method selects that part of the blocks, such as the

green blocks of Figure 4-(c).

As a result, the proposed method uses only 20%

of all HOG blocks. Hence, the overall computational

time is 5 times faster than methods using all the

blocks. And the performance is almost the same.

3 PROPOSED ALGORITHM

The HOG method of pedestrian detection divides an

image into units of cells after calculating orientation

gradients, such as in Figure 5-(b) and then calculates

a histogram, as shown in Figure 5-(e). The oriented

gradients is divided cell such as Figure 5-(c) and

mixed 4 cell to block. Then the block is normalized

by the L1-norm or L2-norm method. In a 48ⅹ96

image, the feature is 1980 because all block number

is 55 and each block is calculated 9 histogram.

However in Figure 5-(c), the non-pedestrian

blocks which correspond to the background or other

objects (for example, blocks 1, 5, 51 and 55) can be

used for extracting the feature. Blocks which do not

contain a pedestrian block, such as those for a

building, tree and road, are not required for

ICINCO2014-11thInternationalConferenceonInformaticsinControl,AutomationandRobotics

784

Figure 6: Hit rate per block (green: high rank 40%, red:

low rank 60%).

Figure 7: Hit rate per block (green: high rank 30%, red:

low rank 70%).

Figure 8: Hit rate per block (green: high rank 20%, red:

low rank 80%).

information processing because the environment

exists in various other shapes, unrelated to

pedestrians.

Accordingly, in this paper, rather than using all

1980 features, only the relevant features are used. As

shown in Figure 6, all of the blocks are each trained

and tested using SVM, and a hit rate is calculated for

each block. Then blocks with a high hit rate are

selected for the feature after the unnecessary data,

having a low hit rate, are removed. In Figure 6, the

green areas are good blocks with a high hit rate. And

these blocks are similar in shape to a human head,

shoulder, waist and leg.

In Figure 6, the green block is high rank 40% of

hit rate such as human. And the red block is low

rank 60% of hit rate such as environment. In Figure

8, the high rank 20% of hit rate similar to pedestrian

head, shoulder, waist and leg. As shown in Figure 8,

in this paper we propose a block selection method.

The proposed method is to analyse each block to

determine their relationship to a pedestrian figure,

then to remove unnecessary blocks such as those

corresponding to the environment, and select good

blocks which correspond to a human head, shoulder,

waist and leg.

4 EXPERIMENTS

This study used the Daimler pedestrian dataset for

pedestrian detection experiments. The Daimler

pedestrian dataset includes 48ⅹ96 images that were

acquired while driving on roads. The training dataset

is composed of 52,112 positive images and 32,465

negative images. The test dataset is composed of

25,608 positive images and 16,235 negative images.

In Figure 9, the test result is the detection rate per

block number. All 55 of block numbers 1 to 55 were

trained and tested.

In Figure 9, the performance is almost the same

as compared with methods using all the blocks. But

the detection rate rapidly decreases when the method

uses less than 11 blocks. The detection rate using all

the blocks is 94.48% and the method using 11

blocks is 93.19%. This means the detection rate

decreases 1.3% when going from using 55 blocks for

detection to using 11 blocks. This is slight change in

performance considering the block number has been

decreased by 80%.

As shown in Figures 10, 11 and 12, we provide

an analysis of the classifier comparing the method

using all blocks and the proposed method. In Figure

10, the classifier result for the method using all

blocks is a false negative in a complex background,

but the proposed method is a true positive. This is

because the proposed method used just a part of the

PedestrianDetectionusingHOG-basedBlockSelection

785

0 5 10 15 20 25 30 35 40 45 50 5

Number of Block

Detect Rate(%)

Figure 9: Detection rate per block number.

Figure 10: Proposed method: true positive; method using

all blocks: false negative.

Figure 11: Proposed method: false negative; method using

all blocks: true positive.

Figure 12: Proposed method: false negative; method using

all blocks: false negative.

HOG blocks. However in Figure 11, the classifier

result is the opposite of the image in Figure 10, with

a simple background. In Figure 12, the method using

all blocks and the proposed method are false

negatives due to overlapping of the pedestrians.

Figure 13: Proposed method: true negative; method using

all blocks: false positive.

Figure 14: Propose method: false positive; method using

all blocks: true negative.

Figure 15: Proposed method: false positive; method using

all blocks: false positive.

Also we analysed the classifier with negative

images. As shown in Figures 13, 14 and 15, the

classifier results are similar to the positive images.

The performance of the proposed method is better

than the method using all blocks in a complex

background. And the method using all blocks and

the proposed method are false positive for large

objects such as a vehicle.

We have experimentally tested the proposed

method in pedestrian detection and found the

performance of pedestrian detection to be good with

persons in a complex background, such as Figure 16.

However vertical or large objects produce a false

positive, as in Figure 17. And overlapping persons

were missed in detection, such as in Figure 18.

5 CONCLUSIONS

In this paper, the proposed algorithm effectively

reduced unnecessary features by feature analysis and

improved computational speed. Existing methods to

ICINCO2014-11thInternationalConferenceonInformaticsinControl,AutomationandRobotics

786

Figure 16: True Positive of pedestrian detection using the

proposed method (black box: detection result).

Figure 17: False Positive of pedestrian detection using the

proposed method (black box: detection result).

Figure 18: False negative of detection using the proposed

method (red box: missed detection).

improve computational speed require either high

complexity or special equipment. But the proposed

algorithm greatly improved computation speed by

using a simple method. The experimental results

show that the computation speed of the proposed

method is 5 times faster than the method using all

HOG blocks. However the performance is almost

the same. The classifier result of the proposed

method is better than the method using all blocks for

complex backgrounds.

Also, pedestrian detection using the proposed

method can provide a real time operating system for

intelligent vehicles. In future works, we will use a

more advanced probability method for partial block

analysis. These studies will be helpful for

developing a system for pedestrian detection for

intelligent vehicles.

ACKNOWLEDGEMENTS

This work was supported by the DGIST R&D

Program of the Ministry of Education, Science and

Technology of Korea.

REFERENCES

Geronimo, D., Lopez, A., M., Sappa, A., D., 2010. Survey

of pedestrian detection for advanced driver assistance

systems. In IEEE Transactions on Pattern Analysis

and Machine Intelligence.

Dalal, N., Triggs, B., 2005. Histogram of Oriented

Gradients for Human Detection. In Proceeding of

IEEE Conference Computer Vision and Pattern

Recognition.

Dollar, P., Wojek, C., Schiele, B., Perona, P., 2012.

Pedestrian detection: an evaluation of the state of the

art. In IEEE Transactions on Pattern Analysis and

Machine Intelligence.

Dollar, P., Belongie, S., Perona, P., 2010. The Fastest

Pedestrian Detector in the West. In Proceeding of

Conference British Machine Vision.

Viola, P., Jones, M., 2001. Fast multi-view face detection.

In Proceeding of IEEE Conference Computer Vision

and Pattern Recognition.

Munder, S., Schnorr, C., Gavrila, D., 2008. Pedestrian

detection and tracking using a mixture of view-based

shape-texture models. In IEEE Transactions on

Intelligent Transportation Systems.

Prisacariu, V., A., Reid, I., 2009. fastHOG – a real-time

GPU implementation of HOG. Technical Report.

PedestrianDetectionusingHOG-basedBlockSelection

787