Utilize este identificador para referenciar este registo:
https://hdl.handle.net/10216/133046Registo completo
| Campo DC | Valor | Idioma |
|---|---|---|
| dc.creator | Luís Filipe Pinheiro Gonçalves Ribeiro | |
| dc.date.accessioned | 2025-11-06T00:38:25Z | - |
| dc.date.available | 2025-11-06T00:38:25Z | - |
| dc.date.issued | 2021-02-12 | |
| dc.date.submitted | 2021-03-15 | |
| dc.identifier.other | sigarra:448924 | |
| dc.identifier.uri | https://hdl.handle.net/10216/133046 | - |
| dc.description | With the time, more and more humanity tries to work together with machines. Now is getting more attention the implementation of machine learning in interaction between machines and humans. This thesis tries to implement an action classification model in industrial environments. Several obstacles were found and some were beaten others were accepted as limitations of the model with an idea of what to do to solve this issues. The whole thesis can be separated in several parts, raw data acquisition, raw data analysis and process, feature extraction, model creation, model testing and improvement, final tests. In the raw acquisition part, we have two different fronts, the data acquisition from the dataset, and the data acquisition from the Zed camera which will be used in the final product. The raw analysis and process studies where is needed refining, such as the number and the position of zeros and nones, studies the person recognized because some frames changes the person recognized to someone in the background instead of always being the one preforming the actions and studies the class balance in the dataset, since although every class has the same number of videos, each video has different lengths which is translated as more samples per class, in classes that have videos with greater length. The analysis part has the purpose to know where to apply process, and the process part is to solve the problems identified in the analysis part. In the Feature extraction, we focused on 3 geometric features since almost the beginning of the thesis. This 3 geometric features are the joint to joint distance, joint to line orientation and distance, line to line distance. Our body model is built from the body25 model, selecting only the 15 first joints, that it's thought that it has the most relevant information of the action. The lines set is a the line between a set of joints that is thought that also have the most relevant information of the action. In the model creation we tried some architectures, the mostly used by us were LSTM but always with negative performances, only in the end, where we tried a CNN architecture we achieve good results. In the model testing and improvement, we trained the model with small differences from one to the next, in order to choose the best architecture. In the final tests, we tested the model with tests from actions performed by us, with a setup very close to the setups with more accuracy. Each setup has different configurations of height and distance of the camera. All this points have illustrations to better explain what is heappening. | |
| dc.language.iso | eng | |
| dc.rights | restrictedAccess | |
| dc.subject | Engenharia electrotécnica, electrónica e informática | |
| dc.subject | Electrical engineering, Electronic engineering, Information engineering | |
| dc.title | Intelligent system for live skeleton-based action recognition | |
| dc.type | Dissertação | |
| dc.contributor.uporto | Faculdade de Engenharia | |
| dc.identifier.doi | 10.34626/frvq-d411 | |
| dc.identifier.tid | 202822583 | |
| dc.subject.fos | Ciências da engenharia e tecnologias::Engenharia electrotécnica, electrónica e informática | |
| dc.subject.fos | Engineering and technology::Electrical engineering, Electronic engineering, Information engineering | |
| thesis.degree.discipline | Mestrado Integrado em Engenharia Electrotécnica e de Computadores | |
| thesis.degree.grantor | Faculdade de Engenharia | |
| thesis.degree.grantor | Universidade do Porto | |
| thesis.degree.level | 1 | |
| Aparece nas coleções: | FEUP - Dissertação | |
Ficheiros deste registo:
| Ficheiro | Descrição | Tamanho | Formato | |
|---|---|---|---|---|
| 448924.pdf Restricted Access | Intelligent system for live skeleton-based action recognition | 9.4 MB | Adobe PDF | Ver/Abrir |
| 448924.1.pdf | Intelligent system for live skeleton-based action recognition | 9.4 MB | Adobe PDF |  Ver/Abrir |
Todos os registos no repositório estão protegidos por leis de copyright, com todos os direitos reservados.