Spanish scientists have developed a system for detecting pistols on images

Pistols are a favorite firearm of many people. They are small, the gun can be hidden under a jacket or in a bag, the accuracy of shooting of many models is very high (of course, provided that the owner has the appropriate skills). Unfortunately, all this applies not only to respectable citizens, military, law enforcement officers, but also to criminals. Small pistols are often used for robbery, kidnapping and murder.

Detect a gun is not as easy as it seems, external signs of their wearing may not be. True, some criminals with firearms do not behave too prudently. Sometimes they show a weapon (by chance or on purpose) somewhere in public, sitting in a car or just outside. If video surveillance systems were equipped with special software to detect such cases, the police would get more information about potential intruders. Perhaps such a system would help make the streets of cities from different countries safer. Such a software platform is not at all fiction, scientists from the University of Granada (Granada, Spain) are currently working on its creation.

The developers of this system believe that if firearms could be detected even before the shot was fired, the crime would be controlled by the police more effectively. The platform is a machine learning service, a neural network, which with high accuracy determines the presence of pistols on images, including real-time video. The service can “see” the gun on substandard videos from YouTube, even in that capacity, if the weapon “lights up” for a quarter of a second.
')
“The level of crimes committed with the use of pistols in different regions of the world is constantly growing,” said Siham Tabik, head of the research team. - “One of the possible ways to reduce the number of such cases is the introduction of a system for the early detection of weapons that will alert law enforcement officers about the danger. In particular, this can be done by equipping such settlements with a video surveillance system in populated areas. ”

Spanish experts have created their own software platform based on the VGG-16 classification model, having trained the system on the ImageNet image database, which contains about 1.28 million photos. In addition, the fine-tuning of the service was carried out using a database of photos of the researchers themselves with 3000 images in it.

Creating a neural network capable of determining the presence of pistols on images of various types is associated with the following problems :

• Every man holds a gun in his own way. Someone holds the weapon with one hand, someone with two;
• The process of creating a sample of images for training is time consuming;
• The system of automatic detection of a pistol in an image should work only when the neural network is “confident” that there really is a weapon;
• The introduction of such a system requires the development of a geolocation module so that, for example, the police receive information about the "glare" of weapons with reference to the site.

The advantage of the new system is that for its final training and improvement of capabilities a small amount of photos with weapons is required. In the case of the development of face detection systems, everything is much more complicated - millions and millions of photographs are required for training, special complex algorithms that evaluate people's faces when searching for possible matches. Here, you only need to evaluate a number of visual parameters that allow you to determine the presence of weapons in the photo.



Among the main points of their work, researchers highlight the following:

• Development of a new database for training systems for early detection of weapons. Experience and data obtained in the course of such work may be useful in other areas;
• Detection of the most universal indicator of detection of pistols on video in real time;
• The introduction of a new criterion, AATpI, which allows assessing the reliability of various weapon detection indicators.

The platform, developed by the Spaniards, can be trained to search and other types of weapons. Now she has all the basic signs characteristic of pistols. This is a large database. For "fine tuning", allowing, to allocate other types of weapons, no longer requires millions of images with new objects, just a few thousand. This approach saves time. The basic principles of such a system can be used to develop other platforms, for example, a system for detecting vehicles of a certain type. The incoming image is evaluated by the software platform for more than 1000 different criteria.

“The neural network showed excellent results even when working with low-quality videos from YouTube. Having processed 30 different scenes, she successfully found a weapon on almost all the scenes where it was, ”the researchers say. In the future, researchers plan to reduce the number of false positives of the platform by introducing new classifiers of objects and refining those that are already in use.