Convolutional Neural Network Model for Weapon Identification

Resumen

This project article a solution to problems with insecurity by detecting the misuse of weapons in public places, it seeks to meet the objectives of creating its own dataset to carry out training and thus be able to evaluate accuracy, precision, sensitivity and F1 Score, validating using different data sets and with these metrics the performance of the model. For the development of the project, the literature related to convolutional neural networks and weapon identification was first investigated, and cameras that allow GPS location to be extracted were also investigated. Then, a dataset made up of network images and own images taken was created. In a real scenario, the model was trained in Google Colab to identify 6 different types of weapons, including firearms and edged weapons. For this, a Yolov8 model was used, which is ideal for detection, classification and segmentation tasks. A program was created in Visual Studio Code to carry out the tests. After the tests were completed, the confusion matrices were created and the different metrics were calculated to evaluate the performance of the model in the different classes. The conclusion was reached after the analysis of the confusion matrices that the model has a medium performance when detecting images of fire (rifles, shotguns, pistols), however it has a low performance when identifying edged weapons (knives, scissors, machetes) especially in knives that are confused with other classes, especially with machetes.