CBL - Campus del Baix Llobregat

Projecte llegit

Títol: Reconocimiento de objetos para el control de drones en el Drone Engineering Ecosystem

Estudiants que han llegit aquest projecte:

Director/a: SALAMÍ SAN JUAN, ESTHER

Departament: DAC

Títol: Reconocimiento de objetos para el control de drones en el Drone Engineering Ecosystem

Data inici oferta: 29-01-2024     Data finalització oferta: 29-09-2024


Estudis d'assignació del projecte:
    GR ENG SIS TELECOMUN
    GR ENG TELEMÀTICA
Tipus: Individual
 
Lloc de realització: EETAC
        Institució/Empresa: Technical University of Catalonia (UPC)
 
Segon director/a (UPC): VALERO GARCÍA, MIGUEL
 
Paraules clau:
Reconocimiento de objetos, dron, video streaming
 
Descripció del contingut i pla d'activitats:
En la actualidad se dispone de una plataforma llamada Drone Engineering Ecosystem (DEE) que permite la gestión de las operaciones de un dron (planificación de misiones, gestión de la información captada por el dron, etc.). La plataforma ha sido desarrollada a partir de las contribuciones de estudiantes de la EETAC, a través de sus TFM y TFG. En el desarrollo de esta plataforma se utilizan tecnologías avanzadas para el desarrollo de aplicaciones web, aplicaciones para móviles y aplicaciones de escritorio (python, vue, ionic, capacitor, git), así como las tecnologías más utilizadas en el mundo de sistemas de drones de código abierto (ardupilot, dronekit, mission planner, etc.). Más información sobre DEE puede encontrarse en https://github.com/dronsEETAC/DroneEngineeringEcosystemDEE

El objetivo inicial de este proyecto es la incorporación de un módulo de reconocimiento de objetos que se utilizarán para guiar la trayectoria del dron. Adicionalmente, se investigarán y evaluarán diferentes técnicas para la transmisión de video a los dispositivos de tierra. Las funcionalidades desarrolladas se pondrán a prueba en el dronLab ubicado en el Campus.


 
Overview (resum en anglès):
The main objective of this final degree project is to develop software that allows a drone to identify and classify objects in real time through the use of a camera. To achieve this, advanced computer vision techniques and artificial intelligence algorithms have been used, ensuring the correct integration between the developed software and the drone control system. A specific use case is to guide the movement of the drone following a route marked by objects strategically located on the ground.
The methodology used to meet the objectives set out focused on the implementation of an object detection module based on the YOLO (You Only Look Once) algorithm, a convolutional neural network optimised for object detection in real time. The module was developed in Python, and its integration into the Drone Engineering Ecosystem (DEE), a drone control platform, enabled the identification of objects and subsequent decision-making by the drone.
During the development process, different YOLOv8 models (v8n, v8s, v8m, v8l, v8x) were selected and evaluated, and then retrained using a proprietary dataset that included classes such as banana, ball, box and backpack. Several tests were performed, both in simulated environments and in a laboratory with a real drone, to measure the accuracy and efficiency of the system. The results were satisfactory, achieving an improvement in object detection compared to pre-trained models, with accuracy increases of up to 53% in some cases.
Despite the achievements, the project had limitations, such as the impossibility of implementing object detection on the drone's Raspberry Pi due to technical problems with the library used, which restricted image processing to the ground equipment. In addition, the resolution of the drone's camera was not optimal for detecting small objects, and some false positives were observed that diverted the drone from its route at times.
In conclusion, the project demonstrated the effectiveness of integrating an advanced object detection system into the DEE, opening the door to future improvements in model accuracy and drone functionality. Future lines of development are suggested to optimise the system, such as the reduction of false positives and the integration of processing on the Raspberry Pi.

© CBLTIC Campus del Baix Llobregat - UPC