| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 57.68 MB | Adobe PDF |
Abstract(s)
Nas últimas décadas, os confrontos militares têm dependido, a uma larga
escala, de novas e avançadas tecnologias, conduzindo a uma mudança de paradigma
ao que respeita as operações militares. Esta mudança deve-se, essencialmente, à tentativa de mitigar erros que se verificaram em operações passadas, otimizar recursos
e salvaguardar vidas humanas, garantindo ao mesmo tempo o sucesso e estado final
desejado de qualquer operação. É neste espectro que são introduzidos os Unmanned
Aerial Vehicles (UAVs) e a significante vantagem que os mesmos possuem em prol
de qualquer operação, desde a redução do custo de vidas humanas, ao apoio direto e
indireto no campo de batalha, ao aumento do conhecimento situacional do terreno e
do inimigo. Neste trabalho, é apresentado como, a partir de uma câmara integrada
de um UAV, é possível fazer a deteção automática de alvos terrestres (militares) e a
sua diferenciação com civis, em tempo real, usando uma rede neuronal de deteção de
objetos. Partindo da aquisição de uma base de dados, grande e variada, capturada
por drones comerciais e imagens extraídas da Internet. Após isto, foi necessário
fazer a escolha da rede neuronal para este fim, que recaiu sobre o YOLOv4-tiny,
sendo uma rede rápida na deteção de objetos em tempo real e com bons resultados
de precisão. De seguida, foi necessário realizar a anotação de todos os militares em
cada imagem, utilizando um programa gratuito de anotação de imagens (DarkLabel
2.4). Após isto, treinámos, validámos e testámos a rede, aferindo quais os melhores
resultados e pesos a serem utilizados na deteção de militares. Por fim, olhámos
para as limitações e trabalhos futuros a serem desenvolvidos na continuidade deste
trabalho
In recent decades, military clashes have depended on a wide range of new and advanced technologies, leading to a paradigm shift in military operations. This change is mainly due to the attempt to mitigate errors that have occurred in past operations, optimize resources and safeguard human lives, while ensuring the success and desired end state of any operation. It is in this spectrum that Unmanned Aerial Vehicles (UAVs) are introduced and their significant advantage in favor of any operation, from the reduction of the cost of human lives, to direct and indirect support on the battlefield, to the increase of situational awareness of the terrain and the enemy. In this paper, it is presented how, from an UAV camera, it is possible to make the automatic detection of ground targets (military) and its differentiation with civilians, in real time, using a neural network of object detection. Starting with the acquisition of a database, large and varied, captured by commercial drones and images extracted from the Internet. After this, it was necessary to choose the neural network, which fell on YOLOv4-tiny, being a fast network in detecting objects in real time and with good accuracy results. Next, it was necessary to perform the annotation of all the military objects in each image, using a free image annotation program (DarkLabel 2.4). After this, we trained, validated and tested the network, to know which results and weights were best to use for military detection. Finally, we looked at the limitations and future work to be done in continuing this work.
In recent decades, military clashes have depended on a wide range of new and advanced technologies, leading to a paradigm shift in military operations. This change is mainly due to the attempt to mitigate errors that have occurred in past operations, optimize resources and safeguard human lives, while ensuring the success and desired end state of any operation. It is in this spectrum that Unmanned Aerial Vehicles (UAVs) are introduced and their significant advantage in favor of any operation, from the reduction of the cost of human lives, to direct and indirect support on the battlefield, to the increase of situational awareness of the terrain and the enemy. In this paper, it is presented how, from an UAV camera, it is possible to make the automatic detection of ground targets (military) and its differentiation with civilians, in real time, using a neural network of object detection. Starting with the acquisition of a database, large and varied, captured by commercial drones and images extracted from the Internet. After this, it was necessary to choose the neural network, which fell on YOLOv4-tiny, being a fast network in detecting objects in real time and with good accuracy results. Next, it was necessary to perform the annotation of all the military objects in each image, using a free image annotation program (DarkLabel 2.4). After this, we trained, validated and tested the network, to know which results and weights were best to use for military detection. Finally, we looked at the limitations and future work to be done in continuing this work.
Description
Keywords
Operações Militares UAVs Deteção de Objetos Redes Neuronais YOLO