| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 12.36 MB | Adobe PDF |
Authors
Advisor(s)
Abstract(s)
Este trabalho teve como objetivo desenvolver e comparar o desempenho de um robô
seguidor de linha utilizando redes neuronais e controladores tradicionais, como o
PID. Após a análise de diferentes técnicas e robôs, foram desenvolvidas e testadas
várias abordagens baseadas em Deep Learning.
Na primeira experiência utilizou-se os sensores infravermelhos do robô para controlador
o robô para seguir uma linha desenhada no chão. Para este controlo, foi
primeiramente implementado um controlador PID. Os dados obtidos com este método
foram depois utilizados para treinar uma rede neuronal LSTM. Esta rede foi
capaz de simular o comportamento do controlador PID e guiar-se pelo percurso
eficazmente.
Na segunda experiência utilizou-se a câmera incluída no robô para realizar a
mesma tarefa de controlar um robô seguidor de linha através de redes neuronais.
Para isto, começou-se por tirar várias imagens do circuito com essa câmera. As
imagens foram depois processadas e posteriormente categorizadas. Para a categorização
foram testados dois métodos, onde o método de categorização manual em 8
diferentes classes mostrou-se ser o mais eficaz para este efeito. Com isto foi treinada
uma rede CNN capaz de categorizar em tempo real novas imagens. Este modelo foi
então utilizado para o controlo do robô.
Os resultados mostraram que as redes neuronais conseguem lidar com trajetos
mais complexos, mas apresentam desafios como maior carga de processamento e
necessidade de calibração. Em contrapartida, o controlador PID se mostrou mais
simples e eficaz em circuitos menos exigentes. Concluiu-se que, as redes neuronais
são promissoras para cenários mais complexos, mas podem ser também utilizadas
eficientemente para um simples seguimento de linha.
The aim of this work was to develop and compare the performance of a line-following robot using neural networks and traditional controllers such as the PID. After analyzing different techniques and robots, several approaches based on Deep Learning were developed and tested. In the first experiment, the robot’s infrared sensors were used to control the robot to follow a line drawn on the floor. For this control, a PID controller was first implemented. The data obtained with this method was then used to train an LSTM neural network. This network was able to simulate the behavior of the PID controller and guide itself along the route effectively. In the second experiment, the camera included in the robot was used to perform the same task of controlling a line-following robot using neural networks. To do this, we began by taking several images of the circuit with the camera. The images were then processed and categorized. For the categorization, two methods were tested, where the manual categorization method into 8 different classes proved to be the most effective for this purpose. This trained a CNN network capable of categorizing new images in real time. This model was then used to control the robot. The results showed that neural networks can handle more complex paths, but present challenges such as a higher processing load and the need for calibration. In contrast, the PID controller proved to be simpler and more effective for less demanding circuits. It was concluded that neural networks are promising for more complex scenarios, but can also be used efficiently for simple line following.
The aim of this work was to develop and compare the performance of a line-following robot using neural networks and traditional controllers such as the PID. After analyzing different techniques and robots, several approaches based on Deep Learning were developed and tested. In the first experiment, the robot’s infrared sensors were used to control the robot to follow a line drawn on the floor. For this control, a PID controller was first implemented. The data obtained with this method was then used to train an LSTM neural network. This network was able to simulate the behavior of the PID controller and guide itself along the route effectively. In the second experiment, the camera included in the robot was used to perform the same task of controlling a line-following robot using neural networks. To do this, we began by taking several images of the circuit with the camera. The images were then processed and categorized. For the categorization, two methods were tested, where the manual categorization method into 8 different classes proved to be the most effective for this purpose. This trained a CNN network capable of categorizing new images in real time. This model was then used to control the robot. The results showed that neural networks can handle more complex paths, but present challenges such as a higher processing load and the need for calibration. In contrast, the PID controller proved to be simpler and more effective for less demanding circuits. It was concluded that neural networks are promising for more complex scenarios, but can also be used efficiently for simple line following.
Description
Keywords
AGV Robô Raspbot PID LSTM CNN IA Deep learning Seguidor de linha Line follower