Hard Negative Mining from in-Vehicle Camera Images based on

Multiple Observations of Background Patterns

Masashi Hontani

, Haruya Kyutoku

, David Wong

, Daisuke Deguchi

Yasutomo Kawanishi

, Ichiro Ide

and Hiroshi Murase

Graduate School of Informatics, Nagoya University, Furo-cho, Chikusa-ku, Nagoya-shi, Aichi, Japan

Information Strategy Ofﬁce, Nagoya University, Furo-cho, Chikusa-ku, Nagoya-shi, Aichi, Japan

{kawanishi, ide, murase}@i.nagoya-u.ac.jp

Keywords:

Pedestrian Detection, Hard Negative Mining, Additional Learning.

Abstract:

In recent years, the demand for highly accurate pedestrian detectors has increased due to the development of

advanced driving support systems. For the training of an accurate pedestrian detector, it is important to collect

a large number of training samples. To support this, this paper proposes a “hard negative” mining method

to automatically extract background images which tend to be erroneously detected as pedestrians. Negative

samples are selected based on the assumption that frequent patterns observed multiple times in the same

location are most likely parts of the background scene. As a result of an evaluation using in-vehicle camera

images captured along the same route, we conﬁrmed that the proposed method can automatically collect false

positive samples accurately. We also conﬁrmed that a highly accurate detector can be constructed using the

additional negative samples.

1 INTRODUCTION

In recent years, there has been a decrease in the num-

ber of trafﬁc accidents because of improvements in

road and vehicle safety. However, even so, in Japan

alone there were about ﬁve hundred thousand trafﬁc

accidents in 2016, and road safety remains an impor-

tant social issue to be overcome (Statistics Bureau,

2017). Therefore, strategies for reducing trafﬁc acci-

dents remain an active area of development. For ex-

ample, in the last few years, technologies for assisting

drivers to drive safely have been put into practical use

in production vehicles.

A fundamental technology for driving assistance

systems is a detection method for pedestrians and ob-

stacles around the vehicle (Redmon et al., 2016; Pre-

mebida et al., 2014). Modern vehicles may employ a

variety of sensors for this purpose—for example, RA-

DAR and cameras. In addition, using advanced sen-

sors such as LIDAR for vehicle applications is an area

of intensive research. However, since a monocular ca-

mera is inexpensive and compact, and can be easily

ﬁtted to existing vehicle designs, this paper aims to

improve the accuracy of pedestrian detectors by em-

ploying a monocular camera.

When constructing a pedestrian detector, it is im-

portant for the classiﬁer to learn the various appearan-

ces of pedestrians. Although in general thousands of

training samples are necessary, manually collecting a

large number of training samples is time consuming

and expensive. Therefore, methods for constructing

a detector from limited annotated samples have been

proposed. In these methods, the initial pedestrian de-

tector is constructed with annotated samples. Then,

training samples are extracted from the detection re-

sults and re-annotated as positive or negative samples.

Finally, a more accurate detector is reconstructed in-

cluding these additional samples. For example, Yuan

et al. proposed a method for constructing highly accu-

rate trafﬁc sign detectors by automatically processing

the results of a base detector (Yuan et al., 2017). Traf-

ﬁc signs from in-vehicle camera images are charac-

teristic in that they have a regular symmetrical shape

and can be stably detected and tracked between se-

quential frames. Therefore, the initial detector’s de-

tection results can be labelled as positive or negative

samples, and used for online learning to improve both

detection and tracking results.

Another method proposed by Mitsugami et al.

constructs a pedestrian detector for ﬁxed cameras,

adapted to capture location (Mitsugami et al., 2013).

In this method, a detection is judged to be false when

Hontani, M., Kyutoku, H., Wong, D., Deguchi, D., Kawanishi, Y., Ide, I. and Murase, H.

Hard Negative Mining from in-Vehicle Camera Images based on Multiple Observations of Background Patterns.

DOI: 10.5220/0007376804350442

In Proceedings of the 14th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2019), pages 435-442

ISBN: 978-989-758-354-4

435

(a) t

(b) t

3HGHVWULDQ

%DFNJURXQG

'\QDPLF

6WDWLF

Figure 1: Three images captured at the same location at different timings, showing detected pedestrians.

the same object is detected at the same position within

sequential frames. Conversely, it is judged to be cor-

rect when the object moves smoothly, adhering to a

pedestrian motion model. Finally, the detector is re-

constructed with the additional information provided

by the positive and negative samples.

These methods show that additional training using

the detection results of the initial detector is useful for

improving the overall detection accuracy. However,

Yuan’s method cannot be applied to detecting dyna-

mic objects since it uses the characteristics of static

objects. Also, Mitsugami’s method assumes that ima-

ges come from a ﬁxed camera, so it cannot be applied

to in-vehicle camera images captured by moving ca-

meras. For driving assistance systems, it is necessary

to automatically extract samples from the detection

results of moving pedestrians from images captured

by a moving, in-vehicle camera.

Meanwhile, hard negative mining is widely ap-

plied for constructing highly accurate object detec-

tors (Felzenszwalb et al., 2010; Dalal et al., 2006;

Shrivastava et al., 2016). In hard negative mining,

false positives that are easily detected as incorrect are

selected and used for re-training the detector. While

negative samples can be supplied by using random

background patches, it is considered efﬁcient to au-

tomatically collect negative samples which caused the

detector to trigger a false alarm, hence the name “hard

negative”.

From the above background, we consider a met-

hod of extracting samples from in-vehicle camera

images and using them for additional training. Be-

cause it is difﬁcult to accurately extract positive sam-

ples without manual annotation, we focus on hard

negative mining by extracting negative samples that

can be reliably extracted using prior scene knowledge.

Therefore, we focus on the difference between the

characteristics of pedestrian regions and background

regions which may trigger a false detection in order to

extract negative samples automatically. An example

where pedestrians are detected in the same location at

a different timing is shown in Fig. 1. As shown here,

pedestrians are dynamic objects which are not typi-

cally captured at the same position over a long period

of time. Conversely, background areas are static. We

propose a method of extracting negative samples con-

sidering this observation.

The contributions of this paper can be summarized

as follows:

1. The proposal of a novel framework of hard nega-

tive mining from non-annotated in-vehicle image

sequences.

2. The combination of spatial and temporal align-

ments for the detection of frequently observed

background patterns.

3. The use of the resulting hard negative samples for

training state-of-the-art deep-learning-based de-

tectors.

The rest of the paper is organized as follows:

Section 2 describes our negative sample selection

method. Then, we report evaluation experiments

using images captured by an in-vehicle camera and

discuss the results in Section 3. Finally, we conclude

this paper in Section 4.

2 HARD NEGATIVE MINING

FRAMEWORK

A pedestrian region detected by a pedestrian detec-

tor can be classiﬁed as either a correct detection (true

positive) if the region truly contains a pedestrian, or a

false detection (false positive) if the region consists of

background. Since we wish to extract only the false

detections as negative samples, the proposed method

collects “hard negative” candidates. Here, a “hard ne-

gative” refers to a falsely detected pedestrian region,

which is then used to create a negative sample, for use

in further training of the pedestrian detector.

False detections made by pedestrian detections

can be of two types (Mitsugami et al., 2013):

1. Detection of background areas with features simi-

lar to a pedestrian.

2. Detection of parts of a pedestrian.

VISAPP 2019 - 14th International Conference on Computer Vision Theory and Applications

436

7HPSRUDODOLJQPHQW

6SDWLDODOLJQPHQW

&DOFXODWHWKHVLPLODULW\

'LVFDUG

6HOHFW

<HV

([WUDFWGHWHFWHGUHJLRQV

6LPLODULWLHV

LQDOO

஼

DUHKLJK

$OLJQPHQW

VWHS

+DUGQ HJDWLYH

PLQLQJVWHS

5HIHUHQFH VHTXHQFH



'DWDEDVHVHTXHQFHV



'DWDEDVHVHTXHQFHV



′

5HIHUHQFH VHTXHQFH



+DUGQHJDWLYH

FDQGLGDWH



'HWHFWHGUHJLRQV



(a) Alignment / hard negative mining steps

3RVLWLYHV DPSOHV

1HJDWLYHVDPSOHV

7UDLQLQJVDPSOHV

7UDLQLQJ

,QLWLDO

GHWHFWRU

5HIHUHQFH

VHTXHQFH



'DWDEDVH

VHTXHQFHV



′

6HOHFWKDUGQHJDWLYHVDPSOHV

+DUGQHJDWLYHVDPSOHV

$GGLWLRQDOWUDLQLQJ

,PSURYHGGHWHFWRU

(b) Reconstruction step

Figure 2: System ﬂow of the proposed method.

Of these two types, the ﬁrst one is likely to occur

in all images captured at a similar spatial location,

as the background area containing a pattern visually

similar to a pedestrian does not change. This may

lead to a high rate of false detections in certain lo-

cations, which is undesirable for a driver assistance

system. For the second type of false detections, alt-

hough the region is incorrectly detected as a complete

pedestrian, in reality, a pedestrian is likely present

nearby—therefore, this type is less obtrusive. Be-

cause of this, and the fact that background regions

tend to remain visually similar over time, in this re-

search, hard negative candidates are extracted from

false detections of the ﬁrst type.

The proposed method prepares multiple image se-

quences captured along the same route. Among them,

the sequence used for detecting negative samples is

called the reference sequence. The other sequences

used for determining whether or not the candidate is

a false detection, are called the database sequences.

The system ﬂow of the proposed method is shown in

Fig. 2. It consists of three steps: the alignment step

(Fig. 2(a)) for associating a reference sequence with

a database sequence, the hard negative mining step

(Fig. 2(a)) for determining whether or not hard ne-

gative candidates are false detections, and the recon-

struction step (Fig. 2(b)) for constructing a detector

using additional training with negative samples.

2.1 Alignment Step

In the alignment step, each of the multiple database

sequences are aligned temporally and spatially in or-

der to compare the reference sequence to the database

sequences.

First, each database sequence is temporally alig-

ned with the reference sequence. It is necessary to as-

sociate each frame within the database sequence with

the frame within the reference sequence that was cap-

tured at the same location. Typically, temporal align-

ment can be performed referring to the visual simila-

rity between the images in the sequences. However,

there are differences of appearance between frames

captured from the same location caused by different

camera trajectories, small pose changes, and occlu-

sion by obstacles. For this reason, it is difﬁcult to

associate frames accurately with a simple image mat-

ching method.

In this research, frames within database and refe-

rence sequences are temporally aligned using a met-

hod of associating frames proposed by Kyutoku et

al. (Kyutoku et al., 2011). This method uses image

features and epipolar geometry to determine a simila-

rity measure between the capture locations of image

frames, allowing temporal alignment to be performed.

Even after image sequences have been temporally

aligned, capture positions and camera poses differ

slightly between aligned frames. This leads to a po-

sitional gap between corresponding pixels of aligned

frames, which makes direct visual comparison of alig-

ned frames for background detection difﬁcult. There-

fore, temporally aligned frames are then aligned spa-

tially using image registration. This is done using

dense optical ﬂow obtained by DeepFlow (Revaud

et al., 2016). An example of spatial alignment of

pixels in a temporally aligned image frame is shown

Hard Negative Mining from in-Vehicle Camera Images based on Multiple Observations of Background Patterns

437

Frame in the reference sequence

Frame in the database sequence

registered by using optical flow

Frame in the database

sequence aligned temporally

Registration

(a) Example of alignment of a frame in the reference se-

quence to a frame in the database sequence

Frame in the database sequence

registered by using optical flow

Frame in the database

sequence aligned temporally

(b) Difference images for the frames in the database se-

quence

Figure 3: Examples of spatial alignment.

in Fig. 3. The image at the top of Fig. 3(a) is the frame

in the reference sequence and the bottom left image is

the frame in the database sequence selected by tempo-

ral alignment. On the bottom right is the image that

has been spatially aligned using DeepFlow. The diffe-

rence images of these frames are shown in Fig. 3(b).

We can see that the gap between the reference and

the database frames is reduced by the spatial align-

ment. This allows us to compare their visual featu-

res at corresponding positions directly in a common

image coordinate system.

2.2 Hard Negative Mining Step

In the hard negative mining step, negative samples are

extracted from hard negative candidates by comparing

the reference sequence with the database sequences.

First, hard negative candidates are detected in the

reference sequence by the initial pedestrian detector.

All positive detection results are considered as hard

negative candidates. In this paper, as one of the state-

of-the-art deep learning based detectors, we employ

YOLOv3 (Redmon et al., 2016) (hereafter referred to

5HIHUHQFH

VHTXHQFH

'DWDEDVH

VHTXHQFHV

([WUDFWGHWHFWHGUHJLRQVDW

WKHVDPHSRVLWLRQDV

WKHKDUGQHJDWLYHFDQGLGDWH





1



2  3

Figure 4: Conceptual diagram of extraction of detected re-

gions.

as the YOLO detector) as a baseline detector.

Next, detected regions D

(n) are extracted from

the database sequences in order to determine whet-

her or not each hard negative candidate C is a false

detection, where n = 1, 2, .., n

MAX

represents the se-

quence number. A conceptual diagram of the ex-

traction of detected areas is shown in Fig. 4. Detected

regions D

(n) are extracted from each corresponding

temporally and spatially aligned database frame at the

same spatial position as the candidate in the reference

frame. The similarities between C from the reference

image and each of D

(n) from the database sequen-

ces are calculated. A CNN model (Ahmed et al.,

2015) is used for calculating the similarity between

C and D

(n). The result is a set of similarity values

(n) (n = 1, 2, ...) for each C.

Finally, it is determined whether or not the candi-

date C is a false detection by using the calculated si-

milarities S

(n). We consider that the corresponding

(n) is high when D

(n) is captured from the same

object or visual region. In this case, C is probably

a false detection from the background, as the region

does not change temporally between sequences. On

the other hand, when the corresponding S

(n) is low,

C is most likely a temporal object such as a pedestrian

that appears in only one of D

(n). It is necessary to

reliably select only C from background areas as false

detections. Therefore, C is determined as a false de-

tection only when the minimum of S

(n) exceeds a

threshold as

min

(n) > T, (1)

in which case the detection is considered a hard ne-

gative sample. A conceptual diagram of hard nega-

tive sample determination using similarity is shown

in Fig. 5.

2.3 Reconstruction Step

In the reconstruction step, a pedestrian detector is re-

trained including hard negative samples.

First, the hard negative candidates judged as a

false positives in the hard negative mining step are

VISAPP 2019 - 14th International Conference on Computer Vision Theory and Applications

438

0DQ\GHWHFWHGUHJLRQV

ZLWKORZVLPLODULW\

'LVFDUG

6HOHFW

6RPHGHWHFWHGUHJLRQV

ZLWKORZVLPLODULW\

6LPLODULWLHVRIDOO

WKHGHWHFWHGUHJLRQV

DUHKLJK

'LVFDUG









࡯





 

 





஼







஼







஼





୑୅ଡ଼





࡯





࡯







Figure 5: Conceptual diagram of hard negative sample selection.

added to the original training samples used to train

the initial detector. The training samples used in con-

structing the initial detector includes annotated pede-

strians as positive samples, and negative samples are

randomly sampled from the background region in the

YOLO freamwork. In the reconsruction step, nega-

tive samples are randomly sampled from the hard ne-

gative samples and added to the initial samples. Fi-

nally, the pedestrian detector is re-trained with these

training samples.

3 EXPERIMENT

Two evaluation experiments using in-vehicle camera

images were conducted in order to conﬁrm the effecti-

veness of the proposed method. Each will be descri-

bed in detail below.

3.1 Experiment on Negative Samples

Selection

An evaluation experiment using multiple in-vehicle

camera images was conducted in order to conﬁrm the

accuracy of selecting negative samples by the propo-

sed method.

3.1.1 Experimental Condition

For this experiment, four sequences composed of

2,000 to 2,600 frames each, were captured by an in-

vehicle camera mounted on a car running along the

same route travelling in the same direction, at diffe-

rent timings throughout the day. A Point Grey Re-

search Grasshopper3 GS3-U3-28S4C-C and a Space

VP-JHF8M-3MP 8 mm telephoto lens were used. Pe-

destrian bounding boxes were manually annotated in

each camera image, resulting in 3,000 to 7,000 pede-

strian frames per sequence.

One of the four in-vehicle camera sequences was

used as the reference sequence and the remaining

three were used as the database sequences. Next, hard

negative candidates were detected from the reference

sequence by the initial pedestrian detector in the refe-

rence sequence. The YOLO pedestrian detector des-

cribed in Sec. 2.2 was trained using the CityPersons

dataset (Zhang et al., 2017; Cordts et al., 2016) and

the Caltech Pedestrian Detection Benchmark Data-

set (Doll´ar et al., 2012; Doll´ar et al., 2009) (hereafter

referred to as the Caltech dataset).

The negative sample selection method was applied

to the hard negative candidates from the reference se-

quence for determining whether or not each candidate

was a false detection. All candidates that were deter-

mined to be false detections were collected as hard

negative samples.

Hard Negative Mining from in-Vehicle Camera Images based on Multiple Observations of Background Patterns

439



















     

&RUUHFWVHOHFWLRQ

,QFRUUHFWVHOHFWLRQ

Figure 6: Extraction performance of negative samples by

the proposed method.

3.1.2 Evaluation Metrics

The purpose of this research is to accurately select

from the initial pedestrian detection results that do not

contain a pedestrian. Therefore, correct and incorrect

selections are deﬁned in this experiment as follows:

• Correct Selection: A detection result that only

contains background is selected as a negative sam-

ple.

• Incorrect Selection: A detection result that con-

tains a pedestrian is incorrectly selected as a ne-

gative sample.

The selected hard negative sample is determined to

be a correct selection when the IoU (Intersection over

Union) for the candidate and the annotated pedestrian

bounding boxes is greater than 0.4. In other cases, it

is determined to be an incorrect selection. Here, we

use the correct and incorrect selection rates with re-

spect to the total number of hard negative candidates

as evaluation metrics.

3.1.3 Results and Discussion

The detected hard negative candidates were compo-

sed of 3,925 pedestrian images and 80,884 back-

ground images. The proposed method was applied to

these hard negative candidates while varying the si-

milarity threshold to judge whether or not each hard

negative candidate is a false detection. As a result of

this, a graph showing the number of correct selecti-

ons and the number of incorrect selections is shown

in Fig. 6. When the number of incorrect selection is

zero, 2,150 background images are selected as hard

negative samples.

When the similarity threshold was set to a value

that yielded no incorrect negative sample selection,

the negative samples were typically background ima-

ges of ﬁve types: tree, utility pole, sign, light, and

(a) Tree (b) Utility pole (c) Trafﬁc sign

(d) Street light (e) Partial sign

Figure 7: Types of the selected hard negative samples.

partial sign, as shown in Fig. 7. We consider that the

types of the selected negative samples were fairly li-

mited due to the relatively consistent environment co-

vered in the dataset. Therefore, in the future, we will

need to conﬁrm that various types of hard negative

samples can be extracted from various driving envi-

ronments.

An example in which spatial alignment failed due

to differences in appearance of the entire image is

shown in Fig. 8. This occurred because frames at dif-

ferent capture locations were sometimes aligned in-

correctly during temporal alignment. If the pixel dis-

tance between an object in two aligned frames is too

large, the optical ﬂow cannot be correctly calculated.

Therefore, it is necessary to improve the accuracy of

temporal alignment between frames in order to avoid

such failed alignments.

Also, an example where the alignment failed due

to the inﬂuence of the shadow is shown in Fig. 9.

The image registration using optical ﬂow tends to fail

when two images contain large difference in appea-

rance. It is necessary to also improve the accuracy of

spatial alignment since such failure can lead to incor-

rect extractions.

3.2 Experiment on Detector

Construction

An evaluation experiment to construct the pedestrian

detector by incorporating the negative samples se-

lected in Sec. 3.1 was conducted in order to conﬁrm

their usefulness for training.

VISAPP 2019 - 14th International Conference on Computer Vision Theory and Applications

440

A frame in the reference sequence

A frame in the database

sequence aligned temporally

A frame in the database

sequence aligned spatially

Figure 8: Example of an

image where spatial alig-

nment failed.

A frame in the reference sequence

A frame in the database

sequence aligned temporally

A frame in the database

sequence aligned spatially

Figure 9: Example of spa-

tial alignment results in-

cluding shadows.

3.2.1 Experimental Condition

The annotated pedestrians of the CityPersons and Cal-

tech training datasets were used as the base training

samples. The negative samples judged as false de-

tections in Sec. 3.1 were also added in the training

samples. The negative samples were selected using

the threshold at which the number of incorrect selecti-

ons was zero (see Fig. 6).

An image sequence (15,445 frames) captured in

a city environment during the day using the same in-

vehicle camera as the dataset used in Sec. 3.1.1 was

used for the evaluation of the pedestrian detectors.

YOLO detectors were constructed using the follo-

wing training methods:

• Proposed: Trained with the CityPersons dataset

and the Caltech dataset up to 150,000 iterations.

Afterward, trained with additional negative sam-

ples selected using the proposed method up to

50,000 iterations.

• Comparative: Trained with the CityPersons da-

taset and the Caltech dataset up to 200,000 itera-

tions. No additional negative samples were inclu-

ded.

Pedestrian detection was then performed on the

evaluation dataset using the proposed and the com-

parative YOLO detectors. The DET (Detection Error

Tradeoff) curve was used as the evaluation matrix.

0.4

0.5

0.6

0.7

0.8

0.0 0.1 0.2 0.3 0.4

Miss rate

False positives per image

Proposed

Comparative

Figure 10: DET curves of each detector.

3.2.2 Result

The DET curves of each detector are shown in Fig. 10.

We conﬁrmed that the accuracy of the detector can be

improved by training with additional negative sam-

ples extracted using the proposed method. For exam-

ple, when we allowed one false positives in ten images

(0.1), the miss rate improved by 4%.

4 CONCLUSION

In this paper, we proposed a method to improve the

accuracy of pedestrian detectors for in-vehicle came-

ras. Speciﬁcally, we focused on the characteristic that

patterns in background regions which are often incor-

rectly extracted as pedestrians are static. Therefore,

negative samples are extracted automatically from

in-vehicle camera image sequences captured multi-

ple times along the same route, by identifying where

false positive patterns are observed repeatedly bet-

ween temporally and spatially aligned images.

In the evaluation experiments, the effectiveness of

the proposed method in extracting negative samples

and the usefulness of the hard negative samples in ad-

ditional training were conﬁrmed.

Future tasks related to this research include impro-

ving the accuracy of temporal and spatial alignment

between image frames, and examining features used

for calculation of the similarity of false positives bet-

ween sequences.

ACKNOWLEDGEMENTS

Parts of this research were supported by MEXT,

Grant-in-Aid for Scientiﬁc Research.

Hard Negative Mining from in-Vehicle Camera Images based on Multiple Observations of Background Patterns

441

REFERENCES

Ahmed, E., Jones, M. J., and Marks, T. K. (2015). An

improved deep learning architecture for person re-

identiﬁcation. In Proc. 2015 IEEE Conf. on Computer

Vision and Pattern Recognition, pages 3908–3916.

Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler,

M., Benenson, R., Franke, U., Roth, S., and Schiele,

B. (2016). The Cityscapes dataset for semantic ur-

ban scene understanding. In Proc. 2016 IEEE Conf.

on Computer Vision and Patt ern Recognition, pages

3213–3223.

Dalal, N., Triggs, B., and Schmid, C. (2006). Human de-

tection using oriented histograms of ﬂow and appea-

rance. In Proc. 9th European Conf. on Computer Vi-

sion, vol. 2, pages 428–441.

Doll´ar, P., Wojek, C., Schiele, B., and Perona, P. (2009). Pe-

destrian detection: A benchmark. In Proc. 2009 IEEE

Conf. on Computer Vision and Pattern Recognition,

pages 304–311.

Doll´ar, P., Wojek, C., Schiele, B., and Perona, P. (2012).

Pedestrian detection: An evaluation of the state of the

art. IEEE Trans. on Pattern Analysis and Machine

Intelligence, 34(4):743–761.

Felzenszwalb, P. F., Girshick, R. B., McAllester, D. A.,

and Ramanan, D. (2010). Object detection with

discriminatively trained part-based models. IEE E

Trans. on Pattern Analysis and Machine Intelligence,

32(9):1627–1645.

Kyutoku, H., Takahashi, T., Mekada, Y., Ide, I., and Murase,

H. (2011). On-road obstacle detection by comparing

present and past in-vehicle camera images. In Proc.

2011 IAPR Conf. on Machine Vision Applications, pa-

ges 357–360.

Mitsugami, I., Hattori, H., and Minoh, M. (2013). Impro-

ving human detection by long-term observation. In

Proc. 2nd IAPR Asi an Conf. on Pattern Recognition,

pages 662–666.

Premebida, C., Carreira, J., Batista, J., and Nunes, U.

(2014). Pedestrian detection combining RGB and

dense LIDAR data. In Proc. 2014 IEEE/RSJ Int. Conf.

on Intelligent Robots and Systems, pages 4112–4117.

Redmon, J., Divvala, S. K., Girshick, R. B., and Farhadi,

A. (2016). You only look once: Uniﬁed, real-time ob-

ject detection. In Proc. 2016 IEEE Conf. on Computer

Vision and Pattern Recognition, pages 779–788.

Revaud, J., Weinzaepfel, P., Harchaoui, Z., and Schmid,

C. (2016). Deepmatching: Hierarchical deforma-

ble dense matching. Int. J. of Computer Vi sion,

120(3):300–323.

Shrivastava, A., Gupta, A., and Girshick, R. B. (2016). Trai-

ning region-based object detectors with online hard

example mining. In Proc. 2016 IEEE Conf. on Com-

puter Vision and Pattern Recognition, pages 761–769.

Statistics Bureau, Ministry of Internal Affairs and Com-

munications, Japan (2017). Japan statistical ye-

arbook 2018. http://www.stat.go.jp/english/data/

nenkan/67nenkan/index.htm (accessed 2018/9/28).

Yuan, Y., Xiong, Z., and Wang, Q. (2017). An incremen-

tal framework for video-based trafﬁc sign detection,

tracking, and recognition. IEEE Trans. on Intelligent

Transportation Systems, 18(7):1918–1929.

Zhang, S., Benenson, R., and Schiele, B. (2017). Cityper-

sons: A diverse dataset for pedestrian detection. In

Proc. 2017 I EEE Conf. on Computer Vision and Pat-

tern Recognition, pages 4457–4465.

VISAPP 2019 - 14th International Conference on Computer Vision Theory and Applications

442