Detekcija hvatanja objekta temeljena na umjetnoj inteligenciji

master's thesis

• Title (english): AI based object grip detection
• Author: Marin Jurković
• Mentor: Matko Orsag (mentor)
• Committee member: Matko Orsag (član povjerenstva)
• Committee member: Frano Petric (član povjerenstva)
• Committee member: Tamara Petrović (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2023-07-14, Croatia

Abstract

U ovom radu obrađeni su problemi semantičke segmentacije i detekcije hvatišta kovčega. Semantička je segmentacija prvotno provedena na sintetičkim podatcima trenirajući neuronsku mrežu, gdje je bilo potrebno zadati parametre i hiperparametre modela koji bi učinili model optimalnim. Postignut je konačni cilj u smislu da se model nije ni prenaučio ni podnaučio, te je pokazao konstantu učenja kroz skoro svih 1000 epoha. Dodatno, s ciljem poboljšanja rezultata učenja stvoreni su dodatni setovi podataka u stvarnom okruženju, s nešto različitim objektima u odnosu na prethodne. Prethodno je mreža trenirana na jednoj vrsti kovčega, dok je dotreniravanja provedeno na stvarnim podatcima koji su uključivali mali kovčeg, veliki kovčeg, te njihovu kombinaciju. Rezultati dotreniravanja donijeli su dodatan napredak, što je pokazalo da je model i s novim tipom podataka nastavio učiti. Zatim je slijedio zadatak detekcije hvatišta kovčega u kojem su implementirana tri algortma. Za svaki algoritam prikazana je implementacija, te zatim rezultati i moguća buduća poboljšanja koja bi naprijedili ove algoritme i učinila ih potencijalnim sastavnim dijelom nekoh drugih problema u području detekcije hvatišta premeta. Algoritmi uparivanja predložaka i pronalaska kontura odbačeni su kao konačna rješenja jer su pokazivali određene slabosti pri generiranju samih točaka hvatišta. Glavni problem je bila raznolikost podataka s smislu promjenjive veličine, orijentacije i pregleda na originalnoj slici. Shodno tome algoritam pronalaska defekta kontura odabran je kao rješenja problema detekcije hvatišta kovčega. Pokazao se kao najrobusiji i izgenerirao je točke hvatišta koje su se u većini slučajeva nalazile na sredini ručaka, što je i bio konačni cilj.

Abstract (english)

This study addresses the problems of semantic segmentation and detection of suitcase handles. Semantic segmentation was initially performed on synthetic data by training a neural network, where it was necessary to set parameters and hyperparameters of the model to achieve it's optimal performance. The ultimate goal was achieved in the sense that the model demonstrated neither overfitting nor underfitting, and it showed constant learning through almost all 1000 epochs. Additionally, with the aim of improving the learning results, additional data sets were generated in the real environment, with slightly different objects compared to the previous ones. Initially, the network was trained on one type of suitcase, while fine-tuning was performed on real data that included a small suitcase, a large suitcase, and their combination. The fine-tuning results brought additional progress, which showed that the model continued to learn even with the new type of data. Then followed the task of detecting the handle of the suitcase, in which three algorithms were implemented. For each algorithm, the implementation is presented, followed by the results and possible future improvements that would advance these algorithms and make them a potential integral part of some other problem in the field of object detection. Algorithms for matching templates and finding contours were rejected as final solutions because they showed certain weaknesses when generating the grip points themselves. The main problem was the variety of data in terms of variable size, orientation and perspective in the original image. Accordingly, the contour defect finding algorithm was chosen as a solution to the problem of detecting the suitcase handle. It proved to be the most robust and generated grip points that in most cases were in the middle of the handle, which was the ultimate goal.

• Keywords: semantička segmentacija; detekcija hvatišta; neuronske mreže; uparivanja predložaka; pronalaženje kontura; pronalaženje defekta kontura; robusnost; točke hvatišta
• Keywords (english): semantic segmentation; handle detection; neural networks; template matching; contour finding; contour defect finding; robustness; handle points
• Language: croatian
• URN:NBN: urn:nbn:hr:168:446394
• Study programme: Title: Information and Communication Technology; Study programme type: university; Study level: graduate; Academic / professional title: magistar/magistra inženjer/inženjerka informacijske i komunikacijske tehnologije (magistar/magistra inženjer/inženjerka informacijske i komunikacijske tehnologije)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2023-10-11 09:13:21