Projecte llegit
Títol: Machine learning on deep neural networks and object tracking applied to motion of airplanes
Estudiants que han llegit aquest projecte:
- MARTIN TORRES, CLAUDIA (data lectura: 14-09-2020)
- Cerca aquest projecte a Bibliotècnica
Director/a: MASSIGNAN, PIETRO ALBERTO
Departament: FIS
Títol: Machine learning on deep neural networks and object tracking applied to motion of airplanes
Data inici oferta: 23-01-2020 Data finalització oferta: 23-09-2020
Estudis d'assignació del projecte:
- GR ENG SIST AEROESP
Tipus: Individual | |
Lloc de realització: EETAC | |
Paraules clau: | |
Object detection, Object tracking, Machine learning | |
Descripció del contingut i pla d'activitats: | |
The objective of the proposed TFG is to develop efficient
algorithms capable of: i) identifying (multiple) objects in static images, and ii) tracking their motion across various frames of a video. The student will start by solving a simple machine learning task (e.g., recognizing hand-written digits), and then proceed to study the more complex case of object recognition in real photos. Finally, he/she will analyze videos containing moving objects (like for example multiple airplanes taxiing around an airport, or acrobatic airplanes performing stunts in the sky), and write an algorithm capable of following their motion. Object recognition and tracking have immediate, fast-growing applications in engineering, telecommunications, informatics, and physics. The director of the proposed TFG is also a guest scientist at the Institute of Photonic Sciences (ICFO), so that if the occasion arises the student will have the chance to visit the labs and the research teams working therein, and attend relevant research seminars. |
|
Overview (resum en anglès): | |
The aim of this project is to understand the concepts underlying machine learning and how to implement those. To achieve this purpose, an exhaustive study of the origins of this technology has been made, describing the most popular types of neural networks, their history, and the architectures and subsequent implementations. Three implementations of neural networks are presented, using world-known datasets. In the last implementation, an exhaustive study has been realized to achieve the best performance algorithm taking into account different settings. In the second part of the project, Detectron2 has been used, an advanced machine learning program that performs object detection. We have worked with this program and executed a study of the motion of moving airplanes, implementing a new method to track objects given a set of images extracted from a given video. |