Publication: Detección de armas en vídeos digitales
Loading...
Official URL
Full text at PDC
Publication Date
2019
Advisors (or tutors)
Editors
Journal Title
Journal ISSN
Volume Title
Publisher
Citations
Exportar
Abstract
Este trabajo ha sido realizado con el propósito de detectar pistolas en vídeos mediante el uso de algoritmos de inteligencia artificial, en concreto, aprendizaje profundo. Está centrado en la detección de pistolas en vídeos, aunque las técnicas usadas se pueden aplicar a cualquier tipo de objetos como drogas, matrículas, caras, personas...
Se ha hecho una amplia investigación previa para observar las técnicas y modelos que forman parte del estado del arte, comparando los resultados de los mismos para saber cuáles son los mejores para las necesidades del trabajo.
Para la obtención de los mejores resultados, se ha decidido hacer 3 experimentos, en los que se han ido mejorando partes del programa. El primer experimento se centra en elegir el modelo que se va a usar, que finalmente es Faster RCNN con la red neuronal base Inception. El segundo experimento es una mejora incremental en la colección de datos, siendo 4 las mejoras. El tercer experimento son cambios en la configuración del modelo, obteniendo así los parámetros que más se ajustan a la detección de pistolas. En este último experimento se hace también una ampliación sobre la colección de datos, así como un balanceo en las imágenes.
Para conseguir este resultado ha sido imprescindible la investigación previa realizada sobre la detección de objetos en el estado del arte. Se ha conseguido mejorar la configuración de un modelo para adaptarlo a la detección de pistolas. Esto ha sido posible gracias al entendimiento de los múltiples parámetros que conforman un modelo.
Una de las grandes aportaciones de este trabajo, que puede pasar desapercibida, es la colección de imágenes que se ha generado gracias al etiquetado manual mediante un programa. Esta colección consta de miles de imágenes de gran relevancia para la detección de pistolas ya sea en vídeos o imágenes.
Se ha conseguido un programa que detecta pistolas en vídeos con una precisión del 90%, una efectividad del 92% y una exactitud del 91%.
This work has been done with the purpose of detecting handguns on videos using artificial intelligence algorithms, or more specific, deep learning. We have focused on handgun detection on videos, even though the techniques used are relevant to any other type of object like drugs, plates, faces, people... There was a previous wide research to observe the techniques and models that are part of the state-of-the-art, comparing their results to know which ones are the better for the needs of this work. For obtaining the best results, there are 5 experiments, where there were improvements on each part of the program. The first experiment was useful to choose the best model for the program, which was the Faster RCNN with the backbone Inception. The second experiment was an incremental improvement of the dataset, being four the number of improvements done in this experiment. The third experiment were changes on the model configuration, obtaining this way the parameters that better fitted to the handgun detection. In this experiment where was also an extension of the dataset. For achieving this result it has been mandatory the initial research that was done over the object detection on the state-of-the-art. There was an improvement in the configuration of the model adapting it to the handgun detection. This has been possible thanks to the understanding of the multiple parameters that are part of the model. One of the biggest accomplishments of this work, that might go unnoticed, is the dataset that was generated thanks to the manual labeling done within a program. This dataset consists of thousends of images of great relevance for handgun detection either on videos or images. It was achieved a program which detects handguns on videos with a precision of 90%, a recall of 92% and an accuracy of 91%.
This work has been done with the purpose of detecting handguns on videos using artificial intelligence algorithms, or more specific, deep learning. We have focused on handgun detection on videos, even though the techniques used are relevant to any other type of object like drugs, plates, faces, people... There was a previous wide research to observe the techniques and models that are part of the state-of-the-art, comparing their results to know which ones are the better for the needs of this work. For obtaining the best results, there are 5 experiments, where there were improvements on each part of the program. The first experiment was useful to choose the best model for the program, which was the Faster RCNN with the backbone Inception. The second experiment was an incremental improvement of the dataset, being four the number of improvements done in this experiment. The third experiment were changes on the model configuration, obtaining this way the parameters that better fitted to the handgun detection. In this experiment where was also an extension of the dataset. For achieving this result it has been mandatory the initial research that was done over the object detection on the state-of-the-art. There was an improvement in the configuration of the model adapting it to the handgun detection. This has been possible thanks to the understanding of the multiple parameters that are part of the model. One of the biggest accomplishments of this work, that might go unnoticed, is the dataset that was generated thanks to the manual labeling done within a program. This dataset consists of thousends of images of great relevance for handgun detection either on videos or images. It was achieved a program which detects handguns on videos with a precision of 90%, a recall of 92% and an accuracy of 91%.
Description
Trabajo de Fin de Grado en Ingeniería Informática, Facultad de Informática UCM, Departamento de Ingeniería del Software e Inteligencia Artificial, Curso 2018/2019.