Computer vision is a branch of arti cial intelligence (AI) whose purpose is giving machine the ability to understand what it "sees" when it is connected to one or more cameras, the fact that make computer vision used for pattern recognition. Current monitoring and control systems still require human monitoring and intervention and their performance depends relatively to human attention. This work presents a system of automatic detection of handguns in videos, suitable for both surveillance and control. We re-formulate this problem of detection in the problem of minimizing false positive detection and solving it using deep convolutional neural network (CNN). This project consists of a study of di erent online handgun detection methods. The method based on supervised deep learning has proved its performance and the results obtained are very encouraging regarding related works despite the technical di culties (material) encountered during the realization of the experiments.
Hamdaoui1 We live in a digital world, where information is stored, processed, indexed and searched using computers, the fact that makes recovery a quick and cheap task. In recent years, considerable progress has been made in the area of object detection. This progress is due the extensive work in this area and the availability of international image databases for machine learning that have allowed researchers to credibly report the simulation of their approaches in this area.
The increase in the use of digital images, motion analysis and object detection in videos has proved that is an indispensable tool for applications as well as video surveillance.
Public safety is a major concern in today's modern society. Weapons creates serious threats to the safety and security of ordinary people, even in the most public places. So dangerous situations in major events may not be avoided by human operators. The question is how can arti cial intelligence based machines be used for total security solution?
The use of online automatic handguns detection can enhance the
performance of surveillance method with a promising application of deep learning.
Recently, this machine learning technique has achieved good performance to
classical techniques such as the naive bayes, decision tree and even with other
types of deep learning algorithms like recurrent neural network in image
classi cation, detection and segmentation [
To overcome these constraints. We proposed to investigate other types of
deep learning model. It consists of using Mobile dedicated model [
The purpose of this paper is to present a system that allows fast and reliable processing of high quality video data and can thus detect and react to the presence of a handgun.
In the rest of this paper, the second section describes and studies motion detection methods and CNNs based model is proposed as well as their interest in the eld of handguns detection. In section three, experimental part of this work is presented with discussion of di erent obtained results. The last section gives a general conclusion about the proposed work. 2
Handgun detection consists in object recognition and nding its position in static
image or in video sequence. In the context of this work, a basic modeling of CNNs
presented in Fig. 1, is proposed and evaluated using the "Handgun Dataset for
the sliding window approach" and "Handgun Dataset for the region proposals
approach" [
The rst step is the acquisition of images. It takes place in a minimum of two stages, taking initial images and taking subsequent images. Motion detection can not be done without the subsequent capture of images. The care taken in taking pictures is crucial for the success of the registration because the quality of the nal results depends on it. 2.1
Unlike in static images, motion detection is an essential process for handguns
detection in a video sequence. As it is a very expensive process in terms of
computation. To avoid unnecessary triggering of handguns detection process in
case of no motion in the video, a few simple image processing operations are
launched to detect the moving object if applicable. [
Di erential images [
A di erential image shows the di erence between two images and makes the movement visible. A di erential image is calculated from three consecutive images It 1, It and It+1. The advantage of this method, is that non-useful background is removed from the result (static information):
I1 = It+1
It; I2 = It
It 1; I =
I1 ^
I2 (2)
Practically, three images are captured at times t-1, t, t+1 to calculate the di erences I1 and I2. I1 returns the absolute di erence between the last two images, while I2 returns the absolute di erence between the rst two images. Finally, the I is calculated between the bits of I1 and I2.
with captured images, it is necessary to convert them to gray level. It is less
complex and more optimal to work with this type of images. On the other hand,
it is necessary to minimize the noise caused by the camera itself and by the
lighting. This is done by averaging each pixel with its neighbors.
Threshold application . In this part of the process, the goal is to convert the
image to binary, that is, to have two possible values. All pixels that exceed the
threshold will be considered as white pixels and other pixels will be considered
as black pixels. This will help to locate the moving object [
Contour detection . Once the region of interest (ROI) in the image is obtained,
the outlines must be detected. At the end of this image, if there is no contour
detected, it is considered that there is no movement and the process returns to
the acquisition step, otherwise it launches the process of handguns detection [
This process is based on a computer vision module that uses a CNN based model
[
The CNN is de ned as a collection of nodes, where a tensor is given as input and another tensor returned at the output of the last nodes. The input tensor is the input image and the output tensor will be the binary classi cation label for detection or non detection.
The detection of handguns addresses several solutions to the problem of
detecting handguns in real time, but the biggest di culty is at hardware level.
This forced us to do the experiments according to three models:
1. CNN based model.
2. Fast R-CNN based model[
3. MobileNet CNN based model[
The rst model presented in g. 2. is composed of ve convolution layers, two pooling layers with maximum output value and three fully connected layers.
The input image is 32 x 32, rstly, the image goes to the rst convolution layer. This layer is composed of 32 lters of size 3 * 3, each of the convolutional layers is followed by a recti ed linear unit (ReLU) as function of activation, this function forces the neurons to return positive values, after this convolution 32 features of size 32 * 32 will be created.
The previously obtained features are sent to the input of the second convolution layer which is also composed of 32 lters, a RELU activation function is applied on the convolution layer, then a pooling is applied to reduce the size. of the image as well as the amount of parameters and calculation. At the exit of this layer, we will have 32 features of size 16 * 16. The same thing is repeated with three, four and ve convolution layers, these layers are composed of 64 lters, the ReLU activation function is always applied on each convolution.
A pooling layer is applied after the ve convolutional layer. At the exit of this layer, we will have 64 features of size 8 * 8. The feature vector resulting from the convolutions has a dimension of 4096.
After these ve convolutional layers, we use a neural network composed of three fully connected layers. The rst two layers each have 1024 neurons where the activation function used is the ReLU, and the third layer is a normalized exponential function that calculates the probability distribution of the 100 classes (number of classes in Handgun Dataset for the sliding window approach). The second model presented in g. 3. receives proposals from regions from an external system (selective search). These proposals will be sent to a pooling layer of the ROI that will resize all regions with their data at a xed size. This step is necessary because the fully connected layer expects all vectors to be the same size. This model has the same architecture as the second model but with some adjustment in the con guration of learning parameters. 3
To evaluate the three proposed models, two databases are used for learning phase.
Handgun Dataset for the sliding window approach. The training data set, suitable for classi cation task, consists of 102 classes with a total of 9261 images. Handgun class at 200.
Handgun Dataset for the region. The training dataset, suitable for detection task, contains 3000 handguns images with a rich context.
Test dataset for classi cation and detection. A total of 608 images including 304 are images of handguns.
Learning of models is done using a sample of 420 images from the Handgun Dataset database for the region proposals approach. The rst model is tested with 608 images (304 handgun images, 304 non-handgun images), the third model with 200 images and the third model with 420 images. 4
The results obtained are improved as much as the network learns the database by increasing the number of learning epochs. Learning database is also a determining element in deep learning, large learning database could help to achieve better results. After analyzing the results obtained, we notice the following:
From Table 1 , Learning error and validation decreases in correspondence with the number of epochs.
For the rst model, we notice that all the misclassi ed images are 274 images, which generate an error rate of 45% and the totality of the well classi ed images is 334 with an accuracy rate of 55%. This unsupervised classi cation model does not give a good result because it needs more learning epochs and the available hardware con guration does not allow this option. On the other hand, for the second model, all the misclassi ed images are 40 images, an error rate of 20% and the totality of the well classi ed images is 160 with an accuracy rate of 80%. This model gives a good result on the static images but for real time test (using webcam), it also needs a very powerful hardware con guration (Computation by GPU, ... etc.). The third model has all misclassi ed images of 42 images, an error rate of 10% and the totality of the well classi ed images is 378, an accuracy rate of 90% for the static images and with a good performance for real time test (see
True Positives (TP): This is the number of images where the process detects a handgun among 304 images that contains a handgun.
True Negative (TN): This is the number of images where the process does not detect a handgun among 304 images that contains a handgun.
False Positive (FP): This is the number of images where the process does not detect a handgun among 304 images that contains a handgun.
False negatives (FN): This is the number of images where the process detects a handgun among 304 images that does not contain a handgun
Accuracy (P): This is the percentage of handgun detection in 304 images that contains handgun.
Recall (R): This is the percentage of handgun detection in the entire 608 images in the test dataset.
F1 measure.
The results obtained are acceptable regarding the results found in [