A Comparison of Techniques based on Image Classiﬁcation and Object

Detection to Count Cars and Non-empty Stalls in Parking Spaces

D. Di Mauro

, A. Furnari

, G. Patan

, S. Battiato

and G. M. Farinella

Department of Mathematics and Computer Science, University of Catania, Catania, Italy

Park Smart s.r.l., Catania, Italy

Keywords:

Counting, Deep Learning, Classiﬁcation, Object Detection, Smart Cities.

Abstract:

The world-wide growth of population in urban areas demands for the development of sustainable technologies

to manage city services, such as transportation, in an efﬁcient way. Motivated by the cost-effectiveness of

image-based solutions, in this paper we investigate the exploitation of techniques based on image classiﬁcation

and object detection to count cars and non-empty stalls in parking areas. The analysis is performed on a dataset

of images collected in a real parking area. Results show that techniques based on image classiﬁcation are very

effective when parking stalls are delimited by marking lines and the geometry of the scene is known in advance.

1 INTRODUCTION AND

MOTIVATIONS

In 2007, as stated in (United Nations, 2008), due to an

unprecedented urban growth, the world’s population

was evenly split between urban and rural areas. As a

result, urban life problems such as air pollution, traf-

ﬁc congestion, and lack of parking spaces, have wor-

sened sharply. Among other possible solutions, the

new concept of Smart Cities has been developed. This

concept proposes the application of recent technolo-

gical advancements in the areas of Internet of Things,

Computer Vision, Machine Learning, 4G Networks,

etc., to improve the liveability of cities. Among the

different applications concerning the concept of smart

cities e.g., trafﬁc analysis (Raymond, 2001; Battiato

et al., 2018), vehicle tracking (Battiato et al., 2015),

etc. in this paper we focus on optimizing the use of

parking spaces in urban areas.

The use of private cars has massively increased

over the years, without a proportional expansion in

the number of parking spaces. Drivers waste a great

deal of time looking for a free parking space and fa-

vor the worsening of trafﬁc and air pollution. Vari-

ous systems aimed at optimizing the use of parking

spaces require methods to count the number of vehi-

cles and free spaces in parking areas. There are three

major types of solutions in this ﬁeld. Counter-based

technologies allow to detect whenever a car enters or

leaves in a closed parking area, equipped with a bar-

rier, in order to update the number of vehicles and the

available parking spaces. Sensor-based technologies

make use of sensors plunged into the asphalt to detect

the presence or absence of cars upon it. Image-based

technologies use cameras monitoring the parking area

and rely on Computer Vision to count vehicles.

Due to the Computer Vision and Machine Le-

arning advances in the last decade, we believe that

image-based solutions are economically advantage-

ous over other methods since they do not require spe-

ciﬁc sensors and can be easily implemented in the

context of free-access parking areas. Nevertheless,

such approaches have to face several variabilities de-

pending on the positioning of the cameras, the shape

of the parking spaces, different lighting conditions,

presence of shadows, occlusions, etc.

In this paper, we investigate the application of two

methods based on Image Classiﬁcation and Object

Detection to tackle two different but related problems:

1) counting the number of empty and non-empty stalls

in a parking area, 2) counting how many cars are pre-

sent in a parking area. The ﬁrst problem is common in

managed parking areas, where parking spots are deli-

mited by lines. The second one arises when there are

not lines delimiting the parking spots and hence the

conﬁguration of the parking lot depends on how dri-

vers park. To perform the analysis, we collected and

labeled images of a parking area in a real scenario.

The rest of the paper is organized as follows.

Section 2 presents the related work. Section 3 ex-

plains the investigated methods. Section 4 introduces

328

Mauro, D., Furnari, A., Patanè, G., Battiato, S. and Farinella, G.

A Comparison of Techniques based on Image Classiﬁcation and Object Detection to Count Cars and Non-empty Stalls in Parking Spaces.

DOI: 10.5220/0006851403280336

In Proceedings of the 15th International Joint Conference on e-Business and Telecommunications (ICETE 2018) - Volume 1: DCNET, ICE-B, OPTICS, SIGMAP and WINSYS, pages 328-336

ISBN: 978-989-758-319-3

the proposed dataset and the experimental settings.

Section 5 reports the results. Section 6 concludes the

paper.

2 RELATED WORKS

We use the verb “counting” to indicate three different

processes which we exploit depending on the num-

ber of instances we need to count (Davis and P

erusse,

1988):

Estimation: the process to approximate the correct

number of elements in a scene. We use this pro-

cess when the number of elements is too big to ex-

tensively count each object instance, e.g. the num-

ber of leaves in a tree;

Subitizing: the process employed when the number

of objects ranges between 1 and 4 instances. The

term has been proposed (Kaufman et al., 1949) to

denote the capability to determine the number of

objects instantly (i.e. at glance); When the number

of objects is higher than 4, the ability to count at

glance decreases in terms of accuracy and speed;

Counting: the process of ﬁnding the exact number of

elements given a ﬁnite set of objects of a speciﬁc

type.

Past neuroscience research has investigated the coun-

ting processes at neural level. Eiselt and Nieder (Ei-

selt and Nieder, 2013) discovered in the primate brain

the presence of rule-selective neurons specialized in

guiding decisions related to a speciﬁc magnitude type

only, as well as generalizing neurons that respond

abstractly to the concept of magnitude rules. These

specialized neurons “number neurons” (Nieder, 2016)

encode the number of elements in a set, as well the

numerosity or cardinality, from both spatial and tem-

poral presentation arrays.

For the Computer Vision community, Object

counting is still a challenging problem which needs a

ﬁne-grained understanding of the scene. The task has

been typically studied considering speciﬁc contexts,

e.g., counting people in crowded scenes (Chan et al.,

2008; Chen et al., 2015; Li et al., 2008; Lempitsky

and Zisserman, 2010; Zhang et al., 2015), cells in

biological images (Lempitsky and Zisserman, 2010),

bacterial colonies (Ferrari et al., 2017), penguins (Ar-

teta et al., 2016), etc.

In particular, counting approaches can be divided

into three groups (Loy et al., 2013):

Counting by Detection: these methods use object

detection and count extensively (Chen et al.,

2015).

Counting by Clustering: these methods assume the

presence of individual entities presenting unique

yet coherent patterns which can be clustered to ap-

proximate the ﬁnal number of instances (Rabaud

and Belongie, 2006).

Counting by Regression: these methods count enti-

ties by learning a direct mapping from low-level

imagery to numbers (Chan et al., 2008; Lempitsky

and Zisserman, 2010; Arteta et al., 2014; Fiaschi

et al., 2012).

The increasing use of surveillance systems is pus-

hing the use of image-based solutions to understand

the semantics of parking areas in order to infer the

presence of free spaces and to count cars. Some

efforts have been done to create datasets useful to

tackle this problem. De Almeida et al. (De Almeida

et al., 2015) built a dataset containing pictures acqui-

red in different climatic conditions (cloudy, sunny,

rainy) and considering three different parking areas.

They have benchmarked the problem of discrimina-

ting empty vs non-empty spaces comparing two diffe-

rent hand-crafted features: Local Binary Patterns and

Local Phase Quantization. The extracted features are

exploited with a SVM classiﬁer.

An approach to analyze events and object trajec-

tories in order to discriminate empty stalls from non-

empty ones is proposed in (Ng and Chua, 2012). The

authors employed motion trajectories as features and

applied the adaptive Gaussian Mixture Model (GMM)

followed by connected component analysis for back-

ground modeling and objects tracking.

Wu et al. (Wu et al., 2007) described a pipeline

in which an image patch is sampled every three stalls

present in a frame. Hand-crafted features are hence

computed to infer the likelihood that a pixel belongs

to ground regions. The features extracted from the

patch representing the three contiguous parking spa-

ces are given to a eight-way Support Vector Machine

(SVM) to classify the 2

possible conﬁgurations of

free or occupied stalls in which the three spaces can

be. Conﬂicts between two neighboring patches are re-

ﬁned by employing a Markov Random Field (MRF).

Deep learning techniques have been recently

adopted in this domain obtaining promising perfor-

mances. In (Amato et al., 2017), a new dataset

(CNRPark-EXT) to train deep learning based models

has been introduced. A modiﬁed AlexNet CNN is

employed to obtain a reduced-size model in order to

make inference possible in real-time on low-cost em-

bedded devices. Di Mauro et al. (Di Mauro et al.,

2016) proposed an evaluation of supervised and semi-

supervised approaches based on a Convolutional Neu-

ral Network (CNN) ﬁne-tuned using respectively la-

beled and pseudo-labeled data. The main aim of the

A Comparison of Techniques based on Image Classiﬁcation and Object Detection to Count Cars and Non-empty Stalls in Parking Spaces

329

Figure 1: An example image acquired in a real parking lot

using a ﬁxed camera with parking stalls highlighted in green

(the image is best seen in digital format). To count non-

empty parking spaces, we assume that the positions of the

stalls are known in advance.

work was to assess which approach is the most con-

venient to balance labeling effort over classiﬁcations

performance.

3 METHODS

As mentioned before, we are interested in counting

the number of empty and non-empty stalls in a par-

king area, as well as in counting how many cars are

present. These two tasks are related but not identical.

The number of cars present in the parking area cannot

be directly inferred from the number of free parking

spaces and the total number of stalls. Indeed, some

vehicles may, for instance, be traversing the parking

spaces, while others might be parked outside the stalls

or they might occupy more than one stall. We assume

a single camera pointing to the parking area, which

acquires images similar to the one shown in Figure 1.

The involved setup is convenient because it reduces

the costs of deployment, trying to monitor all the par-

king lot with the minimum number of cameras.

3.1 Counting Non-empty Spaces

This task consists in counting the number of occu-

pied parking spaces in the monitored area. We assume

that the position of each stall is known in advance (Fi-

gure 1). It should be noted that, since the camera is

ﬁxed, labeling the position of the stalls is part of a

calibration process which needs to be performed only

once (i.e., when the camera is installed). We consider

two approaches to count non-empty parking spaces:

an approach based on image classiﬁcation, and an ap-

proach based on object detection.

Image Classiﬁcation. The ﬁrst approach considers

the problem as an image-based binary classiﬁcation

task. For each stall, we ﬁrst extract the smallest

square image patch containing it. Each image patch is

labeled as “empty” or “full” depending on the occu-

pancy status of the related stall. A classiﬁer is hence

trained to discriminate between “empty” and “full”

stalls. At inference time, the trained classiﬁer is used

to determine the status of each stall in order to obtain

the number of non-empty parking spaces.

Object Detection. The second approach employs

a car detector to localize all the cars present in the

image. All bounding boxes detected with a score lo-

wer than a given threshold d

are discarded. The In-

tersection Over Union (IoU) measure between each

stall and each retained bounding box is hence com-

puted. A stall is deemed to be occupied if the IoU

with at least one detected car is higher than a given

threshold d

. The method allows to count the num-

ber of non-empty parking spaces by determining the

status of each stall. This approach allows to obtain

also information about cars which are parked on non-

marked spaces. Such information can be useful to al-

low for better management of parking areas, e.g. de-

tecting misparked cars.

3.2 Counting Cars

The question we want to address is the following:

“How many cars are present in a given Region of Inte-

rest (RoI)?” This problem focuses on a more general

scenario which does not assume the geometry of the

scene to be known in advance or delimited parking

stalls to be present. Also in this case, we consider

two approaches, one based on image classiﬁcation,

and the other one based on object detection.

Image Classiﬁcation. The ﬁrst approach uses the

output of the binary Image Classiﬁcation described

in the previous section to approximate the number of

cars present in the scene as the number of occupied

stalls. It should be noted that this method requires the

geometry of the scene to be known. Moreover, this

approach cannot deal with cars which are not placed

in any of the marked stalls.

Object Detection. This method uses a vehicle de-

tector to ﬁnd cars present in the scene. Bounding

boxes with detection score lower than the threshold

or with IoU score with respect to the given Region

of Interest lower than a threshold r

are discarded. We

obtain the total number of cars present in the RoI by

considering all retained bounding boxes.

SIGMAP 2018 - International Conference on Signal Processing and Multimedia Applications

330

Table 1: Videos contained in the dataset, along with the

corresponding number of labeled frames.

Camera Video Frame Number

Camera 1

Video 1.1 1,801

Video 1.2 1,801

Camera 2

Video 2.1 1,801

Video 2.2 2,241

Camera 3

Video 3.1 1,321

Video 3.2 2,101

Total 11,066

4 DATASET AND

EXPERIMENTAL SETTINGS

For the purpose of this study, we collected a dataset

comprising a total of 11, 066 images captured during

2 days in our living lab which is located at the campus

of the University of Catania (Figure 2). Each frame

has been labeled with annotations in the form of boun-

ding boxes and parking space conﬁgurations.

The dataset has been acquired using three Full-

HD cameras looking at different parking spaces. The

three cameras are referred to as “Camera 1”, “Ca-

mera 2” and “Camera 3”. “Camera 1” observes 12

parking spaces (Figure 3), “Camera 2” monitors 14

parking spaces (Figure 4), and “Camera 3” acquires

images of 12 parking spaces (Figure 5). Given the dif-

ferent viewpoints of the cameras, the acquired scenes

are characterized by different scene geometries. We

recorded two long videos per camera at 1 f ps. The

two videos have been acquired in different days and

care has been taken to make sure covering many as

possible conﬁgurations of the parking spaces, inclu-

ding the cases in which the parking area was empty

and fully occupied. Table 1 summarizes the videos

contained in the dataset and reports the number of fra-

mes of the considered videos.

For each image frame contained in the dataset, we

labeled a Region of Interest specifying the area within

which the parking spaces are comprised. Each image

has been manually labeled to report:

• the total number of cars present in the monitored

parking space;

• a bounding box around each car inside the moni-

tored parking space;

• a binary vector, each component of which repre-

sents the status of the i-th parking space as empty

(0) or non-empty (1);

• the coordinates of the four corners for each stall

present in the frame (see Figure 1).

We propose two different ways of splitting the

data into training and testing sets. The ﬁrst split as-

sumes that training and testing data have been acqui-

red using a single camera. This gives rise to 6 dif-

ferent data subsets (one for each camera), where one

of the two videos is used for training, and the other

one is used for testing (subset “Nx” in Table 2). The

6 subsets are intended to assess the performance of

methods when exposed to data acquired from a single

camera. The second data split assumes that both trai-

ning and test data have been acquired using the three

cameras. In this case, we obtain two subsets (subset

“X” in Table 2) where data acquired using the three

cameras, but belonging to one of the two videos is

used for training, while the remaining is used for test.

These two subsets are intended to assess the ability of

methods to generalize to different scenes.

All experiments have been performed using the

Caffe library (Jia et al., 2014) on a machine equip-

ped with four NVIDIA GeForce TITAN X with 12Gb

of DDR5 RAM.

4.1 Evaluation Measures

To assess the discrepancy between predicted counts

and ground truth counts, we evaluate the investigated

approaches by computing the Absolute Errors (AE).

Given a test frame I

, the predicted count ˆy

, and the

ground truth count y

, we compute the absolute error

corresponding to I

as follows:

= |y

− ˆy

| (1)

To evaluate the performance on a set of test frames

I = {I

}

i=1

, we also compute statistics of the AE va-

lues computed for each frame, including minimum,

maximum, median and mean. In particular, we consi-

der the standard Mean Absolute Error (MAE), which

is given by:

MAE(I ) =

∑

| ˆy

− y

| (2)

It should be noted that the absolute errors and the de-

rived statistics are easy to interpret, as they are ex-

pressed with the same unit measure of the original

data. E.g., a method reporting a MAE equal to 1 is, in

average, overestimating or underestimating the count

by 1 unit.

To evaluate the performance of the different com-

ponents employed in the investigated methods (i.e.,

image classiﬁcation and object detection), we use the

most appropriate measures. Speciﬁcally, we evaluate

binary image classiﬁcation using accuracy (fraction

of correctly classiﬁed images), whereas object de-

tection using mean Average Precision (mAP) as pro-

posed in (Everingham et al., 2010).

A Comparison of Techniques based on Image Classiﬁcation and Object Detection to Count Cars and Non-empty Stalls in Parking Spaces

331

Figure 2: A satellite image (Google, 2018) of the monitored parking area located at the campus of the University of Catania.

Figure 3: Camera 1 observes 12 par-

king spaces.

Figure 4: Camera 2 observes 14 par-

king spaces.

Figure 5: Camera 3 observes 12 par-

king spaces.

Table 2: Data subsets arising from the two considered data splits and performance of the binary stall classiﬁer, using accuracy

(fraction of correctly classiﬁed stalls); performance of the car detector is measured using standard mean Average Precision

(mAP).

Subset Training Data Testing Data VGG16 Accuracy FasterRCNN mAP

Subset 1a Video 1.1 Video 1.2 0.991 0.224

Subset 1b Video 1.2 Video 1.1 0.987 0.381

Subset 2a Video 2.1 Video 2.2 0.986 0.358

Subset 2b Video 2.2 Video 2.1 0.988 0.185

Subset 3a Video 3.1 Video 3.2 0.949 0.228

Subset 3b Video 3.2 Video 3.1 0.989 0.340

Subset A Videos 1.1, 2.1, 3.1 Videos 1.2, 2.2, 3.2 0.911 0.551

Subset B Videos 1.2, 2.2, 3.2 Videos 1.1, 2.1, 3.1 0.952 0.698

4.2 Image Classiﬁcation

Training and testing data for the binary classiﬁcation

component are obtained by extracting an image pa-

tch around each labeled stall. This is repeated for

each frame of the dataset. The extracted image pat-

ches are hence assigned a binary label depending on

the occupancy status of the stall: 0 for “empty” and

1 for “full”. Using this procedure, we obtain a to-

tal of 17, 688 samples for Video 1.1 (10325 occupied,

7363 empty), 17, 712 samples for Video 1.2 (8463

occupied, 9249 empty), 20, 636 samples for Video 2.1

SIGMAP 2018 - International Conference on Signal Processing and Multimedia Applications

332

Figure 6: Region of Interest (RoI) for

Camera 1.

Figure 7: Region of Interest (RoI) for

Camera 2.

Figure 8: Region of Interest (RoI) for

Camera 3.

(10530 occupied, 10106 empty), 25, 676 samples for

Video 2.2 (12310 occupied, 13366 empty), 13, 032

samples for Video 3.1 (5911 occupied, 7121 empty),

20, 556 samples for Video 3.2 (11304 occupied, 9252

empty).

The binary image classiﬁcation component to discri-

minate between “empty” and “full” stalls has been

implemented by ﬁne-tuning a VGG16 Convolutional

Nerual Network (CNN) (Simonyan and Zisserman,

2014) pre-trained on ImageNet (Russakovsky et al.,

2015). The ﬁne-tuning process is carried out for 10

epochs. A different model is trained on each data sub-

set. Table 2 reports accuracy values on the test sets for

each of the considered data subsets.

4.3 Object Detection

To implement a car detector, we ﬁne-tuned the Fas-

terRCNN (Ren et al., 2015) object detector based on

VGG16 starting from weights pre-trained on Image-

Net. The training process is carried out for 70, 000

iterations using a batch size of 1. As for the detection,

we train a separate model for each data subset. Ta-

ble 2 reports mAP values on the test sets for each of

the considered data subsets.

Threshold Selection. Counting using Object Dete-

cion as discussed in Section 3 make use of two dif-

ferent thresholds to set parking stalls as “empty” or

“occupied” and to count cars. We set such thresholds

to the values which optimize the performance of the

considered method on a validation set which is for-

med randomly selecting 15% of the training samples.

The search for optimal values is performed indepen-

dently on each data subset.

When counting non-empty parking spaces we choose

the values of d

and d

which minimize the MAE on

the validation set. When counting cars we choose the

values of r

and r

which minimize the MAE on the

validation set.

Figure 9: Box plots for counting non-empty spaces. We

plot the mean absolute error for counting non-empty spaces

in the single camera experiment and in the multiple camera.

Higher is the worst.

4.4 RoI Selection

As stated in Section 3.2, we are interested in coun-

ting the number of cars present in a Region of Interest

(RoI) of the frame. For each camera, we annotated a

RoI corresponding with the area including the stalls.

All methods have been tested considering only the se-

lected RoI, hence discarding any detection result not

included in the RoI. Figures 6, 7 and 8 show the RoIs

considered in our experiments.

5 RESULTS

Counting Parking Spaces Results. Table 3 reports

some statistics of the AE values computed for the dif-

ferent experiments performed on the considered data

subsets. Speciﬁcally, the table reports the minimum,

maximum, mean and median Absolute Error over the

considered subset. We also report results for the ag-

gregation of different data subsets. In particular, we

A Comparison of Techniques based on Image Classiﬁcation and Object Detection to Count Cars and Non-empty Stalls in Parking Spaces

333

Figure 11: An example of discriminating empty and non-empty stalls with considered methods.

consider the aggregation of subsets 1a to 3b, each of

which contains images acquired by a single camera,

as well as the aggregation of Subsets A and B, which

contain images acquired by different cameras. Fi-

gure 9 also reports box plots for the AE values con-

tained in the two aggregated sets discussed above.

The best performance is achieved by the method

based on Image Classiﬁcation. This method always

obtains a minimum AE equal to 0 and maximum AE

values not exceeding 5 units. Median errors are often

close to zero. The Mean Absolute Errors of Image

Classiﬁcation methods are overall signiﬁcantly lower

than methods based on Object Detection both in the

case of single camera tests (Subsets 1a to 3b) and

multiple camera tests (Subsets A and B). This obser-

vation is made particularly clear by Figure 9, in which

the box plots related to the Image Classiﬁcation met-

hod exhibit median values and quartile positions lo-

wer than Object Detection. Interestingly, the method

based on Image Classiﬁcation beneﬁts from the pre-

sence of different geometries in the training set, allo-

wing to further lower the MAE of 1.63 to 0.68 (com-

pare “Subsets: 1a, 1b, 2a, 2b, 3a, 3b” to “Subsets: A,

B” in Table 3 and compare left and right plots in Fi-

gure 9). On the contrary, the method based on Object

Detection achieve much worse results. This observa-

tion suggests that the method based on Image Classi-

ﬁcation is more capable to generalize despite can be

used only when stalls are marked. Figure 11 shows a

visual example of the counting stalls results.

Counting Cars Results. Table 4 and Figure 10 re-

port the statistics of AE values and box plots related

to experiments on counting cars. It should be noted

that, in this case, the method based on Image Classi-

ﬁcation should be considered as a baseline, since it is

affected by a systematic error due to the fact that the

number of non-empty spaces is not always equal to

the number of cars (i.e., the method cannot count cars

parked outside the stalls). Nevertheless, similarly to

the previous experiment, the method based on Image

Classiﬁcation achieves the best results (MAE equal to

1.57 in the case of a single camera geometry and 1.16

for multiple camera geometries). This suggest that,

Figure 10: Box plots for counting cars. We plot the mean

absolute error for counting cars in the single camera expe-

riment and in the multiple camera. Higher is the worst

when stalls are present and their location is known,

this information can be exploited to count cars. Inte-

restingly, all the methods beneﬁt from the presence of

different camera geometries in this experiment (com-

pare the two plots in Figure 10).

6 CONCLUSIONS

This paper investigated and compared two different

approaches to count non-empty spaces and cars in

parking areas. To perform the analysis, a dataset

of videos has been collected in a real scenario and

each frame has been labeled according to the position

of parking stalls, the number of occupied stalls, and

the number of cars in the frame. Results show that,

when the geometry of the scene is known (i.e., stalls

are marked), the system can take advantage of binary

classiﬁcation methods to obtain competitive results.

SIGMAP 2018 - International Conference on Signal Processing and Multimedia Applications

334

Table 3: Absolute Errors for the task of counting non-empty stalls. Best results for each data subset are reported in bold

numbers.

Absolute Errors

Subset Method Min Max Mean (MAE) Median

Subsets: 1a, 1b, 2a

2b, 3a, 3b

Image Classiﬁcation 0.00 5.00 1.63 1.00

Object Detection 0.00 7.00 2.15 2.00

Subset 1a

Image Classiﬁcation 0.00 3.00 1.48 1.00

Object Detection 0.00 7.00 4.20 4.00

Subset 1b

Image Classiﬁcation 0.00 2.00 0.54 0.00

Object Detection 0.00 5.00 2.26 3.00

Subset 2a

Image Classiﬁcation 0.00 3.00 2.49 3.00

Object Detection 0.00 5.00 1.54 1.00

Subset 2b

Image Classiﬁcation 0.00 5.00 3.22 4.00

Object Detection 0.00 4.00 2.04 2.00

Subset 3a

Image Classiﬁcation 0.00 2.00 0.88 1.00

Object Detection 0.00 4.00 1.28 1.00

Subset 3b

Image Classiﬁcation 0.00 3.00 0.86 1.00

Object Detection 0.00 3.00 1.82 2.00

Subsets: A, B

Image Classiﬁcation 0.00 2.00 0.56 0.00

Object Detection 0.00 11.00 4.23 4.00

Subset A

Image Classiﬁcation 0.00 1.00 0.40 0.00

Object Detection 0.00 8.00 4.09 4.00

Subset B

Image Classiﬁcation 0.00 2.00 0.68 0.00

Object Detection 0.00 11.00 4.34 4.00

Table 4: Absolute Errors for the task of counting cars. Best results for each data subset are reported in bold numbers.

Absolute Errors

Subset Method Min Max Mean (MAE) Median

Subsets: 1a, 1b, 2a

2b, 3a, 3b

Image Classiﬁcation 0.00 6.00 1.57 1.00

Object Detection 0.00 14.00 2.83 2.00

Subset 1a

Image Classiﬁcation 0.00 6.00 1.44 1.00

Object Detection 0.00 11.00 2.38 2.00

Subset 1b

Image Classiﬁcation 0.00 4.00 0.56 0.00

Object Detection 0.00 11.00 3.43 3.00

Subset 2a

Image Classiﬁcation 0.00 6.00 2.37 2.00

Object Detection 0.00 10.00 2.30 2.00

Subset 2b

Image Classiﬁcation 0.00 6.00 3.15 4.00

Object Detection 0.00 14.00 5.03 5.00

Subset 3a

Image Classiﬁcation 0.00 5.00 0.85 1.00

Object Detection 0.00 7.00 1.92 2.00

Subset 3b

Image Classiﬁcation 0.00 2.00 0.74 1.00

Object Detection 0.00 9.00 1.98 1.00

Subsets: A, B

Image Classiﬁcation 0.00 7.00 1.16 1.00

Object Detection 0.00 11.00 1.95 2.00

Subset A

Image Classiﬁcation 0.00 5.00 0.87 1.00

Object Detection 0.00 9.00 1.77 1.00

Subset B

Image Classiﬁcation 0.00 7.00 1.39 1.00

Object Detection 0.00 11.00 2.10 2.00

A Comparison of Techniques based on Image Classiﬁcation and Object Detection to Count Cars and Non-empty Stalls in Parking Spaces

335

REFERENCES

Amato, G., Carrara, F., Falchi, F., Gennaro, C., Meghini, C.,

and Vairo, C. (2017). Deep learning for decentralized

parking lot occupancy detection. Expert Systems with

Applications, 72(15):327–334.

Arteta, C., Lempitsky, V., Noble, J. A., and Zisserman, A.

(2014). Interactive Object Counting. In European

Conference on Computer Vision, pages 1–15.

Arteta, C., Lempitsky, V., and Zisserman, A. (2016). Coun-

ting in the wild. In European Conference on Computer

Vision, pages 483–498.

Battiato, S., Farinella, G. M., Furnari, A., Puglisi, G., Sni-

jders, A., and Spiekstra, J. (2015). An integrated sy-

stem for vehicle tracking and classiﬁcation. Expert

Systems with Applications, 42(21):7263–7275.

Battiato, S., Farinella, G. M., Gallo, G., and Giudice, O.

(2018). On-board monitoring system for road trafﬁc

safety analysis. Computers in Industry, 98:208 – 217.

Chan, A. B., Liang, Z. S.-J., and Vasconcelos, N. (2008).

Privacy preserving crowd monitoring: Counting pe-

ople without people models or tracking. In Confe-

rence on Computer Vision and Pattern Recognition,

pages 1–7. IEEE.

Chen, S., Fern, A., and Todorovic, S. (2015). Person count

localization in videos from noisy foreground and de-

tections. In Conference on Computer Vision and Pat-

tern Recognition, pages 1364–1372. IEEE.

Davis, H. and P

erusse, R. (1988). Numerical competence in

animals: Deﬁnitional issues, current evidence, and a

new research agenda. Behavioral and Brain Sciences,

11(4):561–579.

De Almeida, P. R., Oliveira, L. S., Britto, A. S., Silva, E. J.,

and Koerich, A. L. (2015). PKLot-A robust dataset

for parking lot classiﬁcation. Expert Systems with Ap-

plications, 42(11):4937–4949.

Di Mauro, D., Battiato, S., Patan

e, G., Leotta, M., Maio,

D., and Farinella, G. M. (2016). Learning approaches

for parking lots classiﬁcation. In International Con-

ference on Advanced Concepts for Intelligent Vision

Systems, pages 410–418. Springer.

Eiselt, A.-K. and Nieder, A. (2013). Representation of ab-

stract quantitative rules applied to spatial and numeri-

cal magnitudes in primate prefrontal cortex. The Jour-

nal of neuroscience: the ofﬁcial journal of the Society

for Neuroscience, 33(17):7526–34.

Everingham, M., Van Gool, L., Williams, C. K., Winn, J.,

and Zisserman, A. (2010). The pascal visual object

classes (voc) challenge. International journal of com-

puter vision, 88(2):303–338.

Ferrari, A., Lombardi, S., and Signoroni, A. (2017). Bacte-

rial colony counting with convolutional neural net-

works in digital microbiology imaging. Pattern Re-

cognition, 61:629–640.

Fiaschi, L., Koethe, U., Nair, R., and Hamprecht, F. A.

(2012). Learning to count with regression forest and

structured labels. In International Conference on Pat-

tern Recognition, pages 2685–2688.

Google (2018). Google Maps. https://www.google.it/maps/

@37.5264537,15.0741852,230m/data=!3m1!1e3?hl=en.

[Online; accessed 12-March-2017].

Jia, Y., Shelhamer, E., Donahue, J., Karayev, S., Long, J.,

Girshick, R., Guadarrama, S., and Darrell, T. (2014).

Caffe: Convolutional architecture for fast feature em-

bedding. arXiv preprint arXiv:1408.5093.

Kaufman, E. L., Lord, M. W., Reese, T. W., and Volkmann,

J. (1949). The discrimination of visual number. The

American journal of psychology, 62(4):498–525.

Lempitsky, V. and Zisserman, A. (2010). Learning to Count

Objects in Images. In Advances in Neural Information

Processing Systems, pages 1324–1332.

Li, M., Zhang, Z., Huang, K., and Tan, T. (2008). Estima-

ting the number of people in crowded scenes by MID

based foreground segmentation and head-shoulder de-

tection. In International Conference on Pattern Re-

cognition, pages 1–4.

Loy, C. C., Chen, K., Gong, S., and Xiang, T. (2013).

Crowd Counting and Proﬁling: Methodology and

Evaluation, pages 347–382. Springer, New York, NY.

Ng, L. L. and Chua, H. S. (2012). Vision-based activities

recognition by trajectory analysis for parking lot sur-

veillance. In International Conference on Circuits and

Systems, pages 137–142.

Nieder, A. (2016). The neuronal code for number. Nature

Reviews Neuroscience, 17(6):366–382.

Rabaud, V. and Belongie, S. (2006). Counting Crowded

Moving Objects. In Conference on Computer Vision

and Pattern Recognition, pages 705–711. IEEE.

Raymond, J.-F. (2001). Trafﬁc analysis: Protocols, attacks,

design issues, and open problems. In Designing Pri-

vacy Enhancing Technologies, pages 10–29.

Ren, S., He, K., Girshick, R., and Sun, J. (2015). Faster R-

CNN: Towards real-time object detection with region

proposal networks. In Advances in neural information

processing systems, pages 91–99.

Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh,

S., Ma, S., Huang, Z., Karpathy, A., Khosla, A.,

Bernstein, M., Berg, A. C., and Fei-Fei, L. (2015).

ImageNet Large Scale Visual Recognition Challenge.

International Journal of Computer Vision (IJCV),

115(3):211–252.

Simonyan, K. and Zisserman, A. (2014). Very deep con-

volutional networks for large-scale image recognition.

arXiv preprint arXiv:1409.1556.

United Nations (2008). World urbanization prospects: The

2007 revision. United Nations Publications.

Wu, Q., Huang, C., Wang, S.-y., Chiu, W.-c., and Chen, T.

(2007). Robust parking space detection considering

inter-space correlation. In International Conference

on Multimedia and Expo, pages 659–662. IEEE.

Zhang, C., Li, H., Wang, X., and Yang, X. (2015). Cross-

scene crowd counting via deep convolutional neural

networks. In Conference on Computer Vision and Pat-

tern Recognition, pages 833–841. IEEE.

SIGMAP 2018 - International Conference on Signal Processing and Multimedia Applications

336