Automatic Handgun Detection Alarm in Videos

Using Deep Learning

arXiv:1702.05147v1 [cs.CV] 16 Feb 2017

Roberto Olmos1 , Siham Tabik1 , and Francisco Herrera1,2

1
Soft Computing and Intelligent Information Systems research
group
2
Department of Computer Science and Artificial Intelligence,
University of Granada, 18071 Granada, Spain. emails:
[email protected], [email protected]

February 20, 2017

Abstract
Current surveillance and control systems still require human super-
vision and intervention. This work presents a novel automatic handgun
detection system in videos appropriate for both, surveillance and con-
trol purposes. We reformulate this detection problem into the problem
of minimizing false positives and solve it by building the key training
data-set guided by the results of a deep Convolutional Neural Networks
(CNN) classifier, then assessing the best classification model under two
approaches, the sliding window approach and region proposal approach.
The most promising results are obtained by Faster R-CNN based model
trained on our new database. The best detector show a high potential even
in low quality youtube videos and provides satisfactory results as auto-
matic alarm system. Among 30 scenes, it successfully activates the alarm
after five successive true positives in less than 0.2 seconds, in 27 scenes.
We also define a new metric, Alarm Activation per Interval (AApI), to
assess the performance of a detection model as an automatic detection
system in videos.
Index terms— Classification, Detection, Deep learning, Convolutional Neu-
ral Networks (CNNs), Faster R-CNN, VGG-16, Alarm Activation per Interval

1 Introduction
The crime rates caused by guns are very concerning in many places in the world,
especially in countries where the possession of guns is legal or was legal for a
period of time. The last statistics reported by the United Nations Office on
Drugs and Crime (UNODC) reveals that the number of crimes involving guns

1
per 100,000 habitants are very high in many countries, e.g., 21.5 in Mexico,
4.7 in United States and 1.6 in Belgium [19]. In addition, several psychological
studies demonstrated that the simple fact of having access to a gun increases
drastically the probability of committing a violent behavior.
One way to reducing this kind of violence is prevention via early detection
so that the security agents or policemen can act. In particular, one innovative
solution to this problem is to equip surveillance or control cameras with an
accurate automatic handgun detection alert system. Related studies address
the detection of guns but only on X-ray or millimetric wave images and only
using traditional machine learning methods [6, 7, 28, 25, 26].
In the last five years, deep learning in general and Convolutional Neural
Networks (CNNs) in particular have achieved superior results to all the classical
machine learning methods in image classification, detection and segmentation
in several applications [18, 13, 22, 8, 29, 23]. Instead of manually selecting
features, deep learning CNNs automatically discover increasingly higher level
features from data [17, 11]. We aim at developing a good gun detector in videos
using CNNs.
A proper training of deep CNNs, which contain millions of parameters, re-
quires very large datasets, in the order of millions of samples, as well as High
Performance Computing (HPC) resources, e.g., multi-processor systems accel-
erated with GPUs. Transfer learning through fine-tuning is becoming a widely
accepted alternative to overcome these constraints. It consists of re-utilizing the
knowledge learnt from one problem to another related one [20]. Applying trans-
fer learning with deep CNNs depends on the similarities between the original
and new problem and also on the size of the new training set.
In general, fine-tuning the entire network, i.e., updating all the weights, is
only used when the new dataset is large enough, else the model could suffer
overfitting especially among the first layers of the network. Since these layers
extract low-level features, e.g., edges and color, they do not change significantly
and can be utilized for several visual recognition tasks. The last layers of the
CNN are gradually adjusted to the particularities of the problem and extract
high level features, which are not readable by the human eye. In this work we
used a VGG-16 based classification model pre-trained on the ImageNet dataset
(around 1.28 million images over 1,000 generic object classes) [24] and fine-tuned
on our own dataset of 3000 images of guns taken in a variety of contexts.
Using CNNs to automatically detect pistols in videos faces several challenges:
• Pistols can be handled with one or two hands in different ways and thus
a large part of the pistol can be occluded.
• The process of designing a new dataset is manual and time consuming.

• The labeled dataset can not be re-utilized by different detection approaches

since they require different preprocessing and labeling operations and can-
not learn from the same labeled databases.
• Automatic pistol detection alarm requires the activation of the alarm in

2
 real time and only when the system is confident about the existence of a
pistol in the scene.
• Automatic detection alarm systems require an accurate location of the
pistol in the monitored scene.

As far as we know, this work presents the first automatic gun detection alarm
system that uses deep CNNs based detection models. We focus on the most used
type of handguns in crimes [30], pistol, which includes, revolver, automatic and
semi-automatic pistols, six-gun shooters, horse pistol and derringers. To guide
the design of the new dataset and to find the best detector we consider the
following steps:

• we reformulate the problem of automatic pistol detection alarm in videos

into the problem of minimizing the number of false positives where pistol
represents the true class and
• we evaluate and compare the VGG-16 based classifier using two different
detection approaches, the sliding window and region proposals approaches.
Due to the particularities of each approach, we applied different optimiza-
tions in each case. We evaluated increasing the number of classes in the sliding
window approach and designing a richer training dataset for the region proposals
approach.
As this work focuses on near real time solutions, we selected the most accu-
rate and fastest detector and assess its performance on seven videos of different
characteristics. Then, we evaluated its suitability as automatic pistol detec-
tion alarm system using a new metric, the Alarm Activation Time per Interval
(AATpI), that measures the activation time for each scene with guns.
The main contributions of this work are:

• Designing a new labeled database that makes the learning model achieve
high detection qualities. Our experience in building the new dataset and
detector can be useful to guide developing the solution of other different
problems.
• Finding the most appropriate CNN-based detector that achieves real-time
pistol detection in videos.
• Introducing a new metric, AATpI, to assess the suitability of the proposed
detector as automatic detection alarm system.
From the experiments we found that the most promising results are ob-
tained by Faster R-CNN based model trained on our new database. The best
performing model shows a high potential even in low quality youtube videos
and provides satisfactory results as automatic alarm system. Among 30 scenes,
it successfully activates the alarm, after five successive true positives, within an
interval of time smaller than 0.2 seconds, in 27 scenes.

3
 This paper is organized as follows. Section 2 gives a brief analysis of the most
related papers. Section 3 provides an overview of the CNN model used in this
work. Section 4 describes the procedure we have used to find the best detector
that reaches good precisions and low false positives rate. Section 5 analyzes
the performance of the built detector using seven videos and introduces a new
metric to assess the performance of the detector as automatic detection system.
Finally the conclusions are summarized in Section 6.

2 Related works
The problem of handgun detection in videos using deep learning is related in
part to two broad research areas. The first addresses gun detection using clas-
sical methods and the second focuses on improving the performance of object
detection using deep CNNs.

2.1 Gun detection

The first and traditional sub-area in gun detection focuses on detecting con-
cealed handguns in X-ray or millimetric wave images. The most representative
application in this context is luggage control in airports. The existent methods
achieve high accuracies by using different combinations of feature extractors
and detectors, either using simple density descriptors [6], border detection and
pattern matching [7] or using more complex methods such as cascade classifiers
with boosting [28]. The effectiveness of these methods made them essential in
some specific places. However, they have several limitations. As these systems
are based on metal detection, they cannot detect non metallic guns. They are
expensive to be used in many places as they require to be combined with X-ray
scanners and Conveyor belts. They are not precise because they react to all
metallic objects.
The second sub-area addresses gun detection in RGB images using classi-
cal methods. The few existent papers essentially apply methods such as SIFT
(Scale-Invariant Feature Transform) y RIFT (Rotation-Invariant Feature Trans-
form), combined with Harris interest point detector or FREAK (Fast Retina
Keypoint) [25, 26, 12]. For example, the authors in [25, 26] developed an ac-
curate software for pistol detection in RGB images. However, their method is
unable to detect multiple pistols in the same scene. The used technique consists
of first, eliminating non related objects to a pistol from the segmented image
using K-mean clustering algorithm then, applying SURF (Speeded Up Robust
Features) method for detecting points of interest. Similarly, the authors in [12]
demonstrated that BoWSS (Bag of Words Surveillance System) algorithm has
a high potential to detect guns. They first extract features using SIFT, cluster
the obtained functions using K-Means clustering and use SVM (Support Vector
Machine) for the training. The authors in [14] addresses riffle detection in RGB
images using SVM (Support Vector Machine).

4
 All the above cited systems are slow, can not be used for constant monitoring,
require the supervision of an operator and can not be used in open areas.

2.2 Detection models

Object detection consists of recognizing the object and finding its location in
the input image. The existing methods address the detection problem by re-
formulating it into a classification problem, they first train the classifier then
during the detection process they run it on a number of areas of the input image
using either the sliding window approach or region proposals approach.

• Sliding window approach: It is an exhaustive method that considers a

large number of candidate windows, in the order of 104 , from the input
image. It scans the input image, at all locations and multiple scales, with
a window and runs the classifier at each one of the windows. The most
relevant works in this context improve the performance of the detection by
building more sophisticated classifiers. The Histogram of Oriented Gra-
dients (HOG) based model [3] uses HOG descriptor for feature extraction
to predict the object class in each window. The Deformable Parts Models
(DPM) [4], which is an extension of HOG based model, uses (1) HOG
descriptor to calculate low-level features, (2) a matching algorithm for de-
formable part-based models that uses the pictorial structures [5] and (3)
a discriminative learning with latent variables (latent SVM). This model
provides very good accuracies for pedestrian detection with a speed of
around 0.07fps and 14s/image.
The obtained accuracies using good classifiers under the sliding window
approach are satisfactory but the detection process can be too slow to be
used in real time.
• Region proposals approach: Instead of considering all the possible win-
dows of the input image as candidates, this approach selects actual can-
didate regions using detection proposal methods [15]. The first detection
model that introduced CNNs under this approach was Region-based CNNs
(R-CNN) [10]. It generates around 2000 potential bounding boxes using
the selective search method [27], warps the obtained regions into images
of the same size then, feeds them to a powerful CNN-based classifier to
extract their features, scores the boxes using SVM, adjusts the bounding
boxes using a linear model, and eliminates duplicate detections via a non-
max suppression. R-CNN provides good performance on the well know
PASCAL-VOC with a speed of 40s/image. Fast R-CNN [9] and subse-
quently Faster R-CNN [21] further improve computation, data access and
disk use of R-CNN. Fast R-CNN has a speed of 0.5 f/s and 2s/image and
Faster R-CNN around 7f/s and 140 ms/image.

This work addresses a new solution to the problem of real-time pistol de-
tection alarm system using deep learning CNN-based detector. We develop,

5
evaluate and compare a CNN based classifier on different new datasets within
the sliding window and region proposals detection based methods.

3 Deep learning model

VGGNet was the first runner-up in ILSVRC 2014 [24]. It was used to show that
the depth of the network is critical to the performance. The largest VGGNet
architecture, VGG-16, involves 144 million parameters from 16 convolutional
layers with very small receptive fields 3 × 3, five max-pooling layers of size 2 × 2
, three fully-connected layers, and a linear layer with Softmax activation in
the output. This model also uses dropout regularization in the fully-connected
layer and applies ReLU activation to all the convolutional layers. This CNN has
a greater number of parameters compared to AlexNet and GoogLeNet, which
makes it more computationally expensive.
Deep CNNs, such as VGG-16, are generally trained based on the prediction
loss minimization. Let x and y be the input images and corresponding output
class labels, the objective of the training is to iteratively minimize the average
loss defined as
N
1 X
J(w) = L(f (w; xi ), yi ) + λR(w) (1)
N i=1
where N is the number of data instances (mini-batch) in every iteration, L is the
loss function, f is the predicted output of the network depending on the current
weights w, and R is the weight decay with the Lagrange multiplier λ. We use
the Stochastic Gradient Descent (SGD), which is commonly used in deep CNNs
to update the weights.

wt+1 = µwt − α∆J(wt ) (2)

where µ is the momentum weight for the current weights wt and α is the learning
rate.
The network weights are randomly initialized if the network is trained from
scratch and are initially set to a pre-trained network weights if fine-tuning the
deep model. In this work we have used fine-tuning VGG-16 and initialized
it with the weights of the same architecture VGG-16 pre-trained on Imagenet
database. The pre-trained VGG-16 model is available through the deep learning
software used in this work, theano with Keras front-end [1, 2] and Caffe [16].

4 Database construction: Towards an equilib-

rium between false positives and false nega-
tives
Automatic pistol detection in videos not only requires minimizing the number
of false positives but also reaching a near real time detection. We analyze the

6
performance of the classifier in combination with two detection methods, the
sliding window (Section 4.1) and the region proposals (Section 4.2).
Due to the differences between these two approaches, different optimization
model based on databases with different characteristics, size and classes, are
applied in each case. In the sliding window approach, we address reducing the
number of false positives by increasing the number of classes and thus building
four databases, Database-1, -2, -3 and -4. The characteristics of all the databases
built in this work are summarized in Table 1. In the region proposals approach,
the detector is directly trained on region proposals of a new database, Database-
5, with richer contexts.

Table 1: Characteristics of the new training- and test-sets. The training sets
are labeled as Database-1, 2, 3 , 4 and 5.

Database- # classes total img # img of pistols # rest of img

1 2 9100 3990 (guns) 5110
2 2 1857 751 1056
3 28 5470 751 4716
4 102 9261 200 9061
5 2 3000 3000 -
Test set 2 608 304 304

To evaluate and compare the performance of the classification model trained

on the five proposed databases, we have built a testset of 608 images distributed
into 304 images that contain pistols and 304 images that do not contain pistols.
We have used three metrics, precision, recall and F1 measure, which evaluates
the balance between the precision and recall. Where
T rue P ositives
precision = ,
T rue P ositives + F alse P ositives
T rue P ositives
recall =
T rue P ositives + F alse N egatives
and
precision × recall
F 1 meaure = 2 ×
precision + recall
For implementing the proposed models we used theano with Keras as front-
end [1, 2] in the sliding window based approach shown in Section 4.1 and
Caffe [16] in the region proposal approach shown in Section 4.2.

4.1 Sliding window approach

This section aims at finding the best combination of classifier and training
dataset within the sliding window detection approach. We address the mini-
mization of the number of false positives by building different training datasets
and increasing the number of classes.

7
Figure 1: Examples from Database-2, the top three images represent the pistol
class and the down three images represent the background class

4.1.1 Two classes

Table 2: The results obtained on the testset by the classification model.

Database- #TP #FN #TN #FP Prec. Rec. F1 meas.

1 32 272 109 191 22.70% 14,35 % 10.53%
2 98 206 293 11 89,91% 32,24% 47,46%
3 85 219 299 5 94,44% 27,96% 43,15%
4 97 207 298 6 94,17% 31,91% 47,67%

To guide our design, and as a first approximation, we have built a preliminary

database, Database-1, and consider a broader class of type gun. Database-1
contains 9100 images extracted from online gun catalogues, gun use tutorials
and gun advertisement (see details in Table 1). We distributed it into 3990
images of class gun, which includes different types of guns, pistols, machine
guns, rifle, machine guns, grenade, multiple rocket launcher, tank and 5110
images of not-gun class, which contains different kind of objects.
We have evaluated the VGG-16 based classification model considering two
classes to indicate either the presence or absence of a pistol in the scene. Table 2
provides the number of true positives, #TP, the number of false positives, #FN,
the number of true negatives, #TN, the number of false positives, #FP, preci-
sion, recall and F1 measure obtained by the classification model when trained
on Database-1, 2, 3 and 4 respectively.
As shown in this Table, the classification model trained on Database-1 ob-
tains a high number of false positives, 191, high number of false negatives, 272,
and consequently low precision and recall. This can be explained by the fact
that the large variety of guns makes the learning task very difficult. To make
the problem affordable, we instead focus our task to pistol detection.
In a second step, we have built Database-2 using 1857 images, 751 images
of class pistol and 1056 of background. We included to the background class

8
different images of hands holding different objects other than pistols, e.g., cell
phone and pen as illustrated in Figure 1. On the test set, the binomial classifi-
cation model trained on Database-2 obtained 11 false positives, a high number
of false negatives 206, a precision of 89, 91%, recall 32, 24% and F1 measure
47, 46%, which are still below our expectations. By analyzing the false positives
we found that most of them consider the white background as part of the pistol
which is due to the presence of the white background in most training examples.

4.1.2 Multiple classes

To further decrease the number of false positives we designed two new databases,
Database-3 and Database-4, and considered a higher number of classes.
Database-3 contains 5470 images set distributed into 28 classes. 751 of
training examples contain pistols. As shown in Table 2, the 28-classes based
classification model overcomes the binomial model by reducing the number of
false positives to 5 and improving the precision to 94, 44%, recall to 27, 96%
and F1 measure to 43, 15%. However, the number of false negatives is still very
high.
We have also explored the performance of the classification model using a
higher number of classes on Database-4. Database-4 is built using 9261 images
distributed into 102 classes, 200 training examples contain pistols and 9061 con-
tain diverse objects, e.g., airplanes, ant, beaver, brain, chair, cell phone, animals
and so on. This classifier produces the best results over the previous ones, a
lower number of false negatives 207, slightly lower number of true negatives 298
and consequently better recall 31, 91% and F1 measure 47, 67%. Next we will
evaluate the detection process using the 102-classes based classifier.

4.1.3 Detection process and final analysis

We selected the best classification model to evaluate it under the sliding window
approach. The classifier is applied automatically, in steps of 60 × 60-pixels, to
windows of 160 × 120-pixels of each input image to determine wether it contains
a pistol. The whole detection process takes 1, 5 seconds in a 640 × 360-pixels
input image. Which is not acceptable for an automatic system such as the one
considered in this work.
The detection model that makes use of the sliding window approach achieves
few false positives and high precision, however, it obtains a low recall 35% and
its execution time is not appropriate for online detection. In next section we
will explore an other alternative to further improve performance and speed of
the detection process.

4.2 Region proposals approach

In this approach, we have used Faster Region based-CNN (Faster R-CNN) [21],
which combines the selective search method with VGG-16 based classifier.

9
Table 3: The results obtained by the classification model under the region pro-
posals approach on the testset.

Database- #TP #FN #TN #FP Prec. Rec. F1 meas.

5 304 0 247 57 84,21% 100.00% 91.43%

Figure 2: Three examples from Database-5. This dataset includes, in addition

to the images, the xml file of each image with the localization information of
the pistols.

The design of a new training dataset for this approach is also manual and can-
not re-use the databases from the previous approach. We have built Database-5
using 3000 images that contains pistols in different contexts and scenarios, down-
loaded from diverse web-sites. Figure 2 provides three examples of Database-5.
We considered a two class model and labeled the pistols by providing its lo-
calization, i.e., bounding box, in each individual training image. The rest of
objects in the image are considered background.
In general, as it can be seen from Table 3, Faster R-CNN trained on Database-
5 obtains the highest performance over all the previously analyzed models. It
provides the highest true positives number and the highest true negative num-
bers and consequently the highest recall 100% and F1 score 91.43%. However
it produces more false positives, 57, and consequently lower precision, 84, 21%.
In Section 5, we address this issue in the context of automatic alarm systems
by activating the alarm only when at least five successive false positives happen
in five successive frames. Next we analyze the speed of this detection model.

4.2.1 Detection process and final analysis

We have evaluated the detection process using Faster-RCNN, which is based
on the region proposals approach. We considered two classes and trained the
classification model on Database 5. The whole detection process in a 1000 ×
1000–pixels image takes 0, 19 seconds approximately producing a rate of 5.3
frames/s. Which allows the pistol detection to be performed in videos in near
real time.
The detection model based on region proposals approach achieves the max-
imum recall 100% over the pistol class, zero false negatives, good precision

10
Table 4: The total number of True Positives #TP, total number of Ground Truth
true Positives #GT P, total number of False Positives #FP in the considered
seven videos, labeled as video 1 to 7

video# #frames #TP #GT P #FP Precision Recall F1 measure

1 393 60 162 8 88.24% 37.04% 52.17%
2 627 467 778 11 98.70% 60.03% 74.36%
3 441 25 58 15 62.50% 43.10% 51.02%
4 591 6 54 0 100.00% 11.11% 20.22%
5 627 24 105 21 53.33% 22.86% 32.00%
6 212 141 290 30 82.46% 48.62% 61.17%
7 501 166 476 6 96.51% 34.87% 51.23%

85.21%, a reasonable false positives number and can be used for real time de-
tection. This makes it a good candidate for detecting pistols in a sequence of
frames as shown in Section 5.

5 Analysis of the detection in videos

In this section we explore the strengths and weaknesses of our model on seven low
quality youtube videos. In particular, we first assess the quality of the detection
and localization (Section 5.1) then analyze the suitability of our model as pistol
detection alarm system using a new metric (Section 5.2).

5.1 Analysis of detection and localization

This section analyzes the performance of our best detection model on seven
public videos, with low qualities, downloaded from youtube. Six of them are
pieces of well known films from the 90s, James Bon: The World is Not Enough
(video 1), 3 pieces from Pulp Fiction (video 2, 3 and 4), Impossible Mission:
Rogue Nation (video 5) and Mister Bin (video 6). The seventh video is a long
pistol threatening video (video 7). The videos with the detections can be found
in a public repository in github 1 .
For the experiments, we analyzed the results of the detection in the videos
frame by frame and consider a detection as true positive if the overlapping
between the handled pistol and the predicted bounding box is more than 50%.
Recall that the way pistols are handled is also a key to the detection. We
consider a pistol as ground truth when it is recognizable by the human eye.
Table 4 provides the total number of True Positives #TP, the total number of
False Positives #FP and total number of Ground Truth Positives #GT P, in
each one of the seven videos. We consider a threshold ∈ [0.7 0.9].
In general, although the scenes are dynamic in most videos, the detector
achieves good balance between precision and recall, especially in videos 2 and
1 https://github.com/SihamTabik/Pistol-Detection-in-Videos.git

11
 Figure 3: An example of an accurate detection of four pistols.

6. See example of an accurate detection in Figure 3. In particular, the detector

provides very high precisions in videos 1, 2, 3, 4 ,6 and 7, and the obtained
number of false positives is very low in all the videos. Which is essential to
avoid activating negative alarms. The obtained false positive detections can be
addressed in a realistic system by activating the alarm only when false positives
are detected in a number of consecutive frames.
The low recall can be explained by the false negatives detected in the frames
with very low contrast and luminosity as shown in Figure 5. The false negatives
depends on the quality of the frame and whether the pistol is clearly visible. In
particular, they occur when the pistol is moved very fast or when it is placed in
the background as in the frame shown in Figure 4.
In conclusion, the obtained results can be considered as acceptable to detect
sequences of clearly visible pistols as it will be analyzed in next section.

5.2 Analyzing the model as alarm detection system

In an automatic pistol detection system the alarm must be activated when the
system is completely confident about the presence of pistols in the scene. To
assess the performance of our detection model as an alarm system, we define
a new metric, Alarm Activation Time per Interval (AATpI). AATpI quantifies
how fast is the system in detecting pistols in a given scene. AATpI is the time
the automatic detection alarm system takes to activate an alarm when it detects
at least k successive frames of true positives. For the analysis presented in this

12
Figure 4: An illustrative example of false negatives, i.e., the two pistols in
background.

Figure 5: Examples of false positives from Video 1.

section we used k = 5.
For the experiments, we selected 30 scenes from the previously used videos
with the next requirements. Each scene is made up of at least 5 frames, filmed
in a fixed scenario, i.e., in the same place, and the pistols are clearly visible to
a human viewer. These scenes can be found in a public repository in github 2 .
The model successfully detects the pistol in 27 scenes with an average time
interval AATpI=0.2 seconds, which is good enough for an alarm system. The
detector fails to detect pistols only in three scenes. This is due to the same
reasons highlighted previously, which are the low contrast and luminosity of the
frames, the pistol is moved very fast or when the pistol is not in the foreground.
In summary, although we have used low quality videos for the evaluation,
the proposed model has shown good performance and demonstrated to be ap-
propriate for automatic pistol detection alarm systems.
2 https://github.com/SihamTabik/Pistol-Detection-in-Videos.git

13
6 Conclusions and future work
This work presented a novel automatic pistol detection system in videos appro-
priate for both, surveillance and control purposes. We reformulate this detection
problem into the problem of minimizing false positives and solve it by building
the key training data-set guided by the results of a VGG-16 based classifier, then
assessing the best classification model under two approaches, the sliding win-
dow approach and region proposal approach. The most promising results have
been obtained with Faster R-CNN based model, trained on our new database,
providing zero false positives, 100% recall, a high number of true negatives and
good precision 84,21%. The best detector has shown a high potential even in
low quality youtube videos and provides very satisfactory results as automatic
alarm system. Among 30 scenes, it successfully activates the alarm after five
successive true positives within an interval of time smaller than 0.2 seconds, in
27 scenes.
As present and future work, we are evaluating reducing the number of false
positives, of Faster R-CNN based detector, by preprocessing the videos, i.e.,
increasing their contrast and luminosity, and also by enriching the training set
with pistols in motion. We will also evaluate different CNNs-based classifier
such as, GoogLenet and consider a higher number of classes.

Acknowledgments
This work was partially supported by the Spanish Ministry of Science and Tech-
nology under the project TIN2014-57251-P. Siham Tabik was supported by the
Ramon y Cajal Programme (RYC-2015-18136).

References
[1] Rami Al-Rfou, Guillaume Alain, Amjad Almahairi, Christof Angermueller,
Dzmitry Bahdanau, Nicolas Ballas, Frédéric Bastien, Justin Bayer, Ana-
toly Belikov, Alexander Belopolsky, et al. Theano: A python frame-
work for fast computation of mathematical expressions. arXiv preprint
arXiv:1605.02688, 2016.
[2] François Chollet. Keras: Theano-based deep learning library. Code:
https://github. com/fchollet. Documentation: http://keras. io, 2015.
[3] Navneet Dalal and Bill Triggs. Histograms of oriented gradients for human
detection. In 2005 IEEE Computer Society Conference on Computer Vision
and Pattern Recognition (CVPR’05), volume 1, pages 886–893. IEEE, 2005.

[4] Pedro F Felzenszwalb, Ross B Girshick, David McAllester, and Deva Ra-
manan. Object detection with discriminatively trained part-based mod-
els. IEEE transactions on pattern analysis and machine intelligence,
32(9):1627–1645, 2010.

14
 [5] Pedro F Felzenszwalb and Daniel P Huttenlocher. Pictorial structures for
object recognition. International Journal of Computer Vision, 61(1):55–79,
2005.
[6] Greg Flitton, Toby P Breckon, and Najla Megherbi. A comparison of
3d interest point descriptors with application to airport baggage object
detection in complex ct imagery. Pattern Recognition, 46(9):2420–2436,
2013.
[7] Richard Gesick, Caner Saritac, and Chih-Cheng Hung. Automatic image
analysis process for the detection of concealed weapons. In Proceedings
of the 5th Annual Workshop on Cyber Security and Information Intelli-
gence Research: Cyber Security and Information Intelligence Challenges
and Strategies, page 20. ACM, 2009.
[8] Mostafa Mehdipour Ghazi, Berrin Yanikoglu, and Erchan Aptoula. Plant
identification using deep neural networks via optimization of transfer learn-
ing parameters. Neurocomputing, 2017, in press.

[9] Ross Girshick. Fast r-cnn. In Proceedings of the IEEE International Con-
ference on Computer Vision, pages 1440–1448, 2015.
[10] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich
feature hierarchies for accurate object detection and semantic segmenta-
tion. In Proceedings of the IEEE conference on computer vision and pattern
recognition, pages 580–587, 2014.
[11] Yanming Guo, Yu Liu, Ard Oerlemans, Songyang Lao, Song Wu, and
Michael S Lew. Deep learning for visual understanding: A review. Neuro-
computing, 187:27–48, 2016.

[12] Nadhir Ben Halima and Osama Hosam. Bag of words based surveillance
system using support vector machines. International Journal of Security
and Its Applications, 10(4):331–346, 2016.
[13] Geoffrey Hinton, Li Deng, Dong Yu, George Dahl, Abdel-rahman Mo-
hamed, Navdeep Jaitly, Andrew Senior, Vincent Vanhoucke, Patrick
Nguyen, Tara Sainath, and Brian Kingsbury. Deep Neural Networks for
Acoustic Modeling in Speech Recognition: The Shared Views of Four Re-
search Groups. IEEE Signal Process. Mag., 29(6):82–97, 2012.
[14] Osama Hosam and Abdulaziz Alraddadi. K-means clustering and support
vector machines approach for detecting fire weapons in cluttered scenes.
Life Science Journal, 11(9), 2014.
[15] Jan Hosang, Rodrigo Benenson, Piotr Dollár, and Bernt Schiele. What
makes for effective detection proposals? IEEE transactions on pattern
analysis and machine intelligence, 38(4):814–830, 2016.

15
[16] Yangqing Jia, Evan Shelhamer, Jeff Donahue, Sergey Karayev, Jonathan
Long, Ross Girshick, Sergio Guadarrama, and Trevor Darrell. Caffe:
Convolutional architecture for fast feature embedding. arXiv preprint
arXiv:1408.5093, 2014.
[17] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet clas-
sification with deep convolutional neural networks. In Advances in neural
information processing systems, pages 1097–1105, 2012.
[18] Quoc V. Le. Building high-level features using large scale unsupervised
learning. In 2013 IEEE Int. Conf. Acoust. Speech Signal Process., pages
8595–8598, 2013.
[19] United Nations Office on Drugs and Crime (UNODC). Global study on
homicide 2013. Data: UNODC Homicide Statistics 2013, 2013.
[20] Sinno Jialin Pan and Qiang Yang. A survey on transfer learning. IEEE
Transactions on knowledge and data engineering, 22(10):1345–1359, 2010.
[21] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn:
Towards real-time object detection with region proposal networks. In Ad-
vances in neural information processing systems, pages 91–99, 2015.
[22] Tara N. Sainath, Abdel-rahman Mohamed, Brian Kingsbury, and Bhuvana
Ramabhadran. Deep convolutional neural networks for LVCSR. In 2013
IEEE Int. Conf. Acoust. Speech Signal Process., pages 8614–8618, 2013.
[23] Xiangbo Shu, Yunfei Cai, Liu Yang, Liyan Zhang, and Jinhui Tang. Com-
putational face reader based on facial attribute estimation. Neurocomput-
ing, 2016, in press.
[24] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks
for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014.
[25] Rohit Kumar Tiwari and Gyanendra K Verma. A computer vision based
framework for visual gun detection using harris interest point detector.
Procedia Computer Science, 54:703–712, 2015.
[26] Rohit Kumar Tiwari and Gyanendra K Verma. A computer vision based
framework for visual gun detection using surf. In Electrical, Electronics,
Signals, Communication and Optimization (EESCO), 2015 International
Conference on, pages 1–5. IEEE, 2015.
[27] Jasper RR Uijlings, Koen EA Van De Sande, Theo Gevers, and Arnold WM
Smeulders. Selective search for object recognition. International journal of
computer vision, 104(2):154–171, 2013.
[28] Zelong Xiao, Xuan Lu, Jiangjiang Yan, Li Wu, and Luyao Ren. Automatic
detection of concealed pistols using passive millimeter wave imaging. In
2015 IEEE International Conference on Imaging Systems and Techniques
(IST), pages 1–4. IEEE, 2015.

16
[29] Wei Yu, Kuiyuan Yang, Hongxun Yao, Xiaoshuai Sun, and Pengfei Xu.
Exploiting the complementary strengths of multi-layer CNN features for
image retrieval. Neurocomputing, 2016, in press.
[30] Marianne W Zawitz. Guns used in crime. Washington, DC: US Department
of Justice: Bureau of Justice Statistics Selected Findings, publication NCJ-
148201, 1995.

17