FER Repository
Not a member yet
12128 research outputs found
Sort by
Preserving Clusters in Synthetic Data Sets Based on Correlations and Distributions
No full textThe rising popularity of machine learning has resulted in quality data becoming increasingly valuable. However, in some cases, the data are too sparse to effectively train an algorithm or the data cannot be disclosed to unaffiliated researchers due to privacy concerns. The sparsity of data may also affect various data analyses that require a certain volume of data to be accurate. One possible solution to the aforementioned problems is data generation. However, to be a viable solution, data generation must simulate real-life data well. To this end, this paper tests whether a previously presented iterative data generation method that generates synthetic data sets based on the attribute distributions and correlations of a real-life data set can faithfully reproduce a clustered data set. The approach is shown to be ineffective for the proposed application, and consequently, a new method is introduced that might preserve the clusters present in the real-life data set. The new method is demonstrated to not only preserve the clusters within the synthetic data set, but also improve the similarity of the attribute correlations of the synthetic data set and the real-life data set
Model Development for Predicting Optimal Inventory for Social Events
No full textOvo istraživanje usmjereno je na analizu potrošnje pića na koncertima, s posebnim
fokusom na utjecaj temperaturnih uvjeta i glazbenog žanra na konzumaciju različitih
vrsta pića. Uvidom u te obrasce, istraživanje nastoji razviti alat koji bi omogućio ugostiteljima
da preciznije procijene količine potrebnih pića za događanja, s ciljem minimiziranja
troškova i optimizacije nabave.
Korištenjem modela strojnog učenja, poput Random Foresta, XGBoost-a i neuronskih
mreža, evaluirala se njihova sposobnost predviđanja temeljem povijesnih podataka
o potrošnji i augmentiranih podataka. Iako savršena predikcija nije moguća, modeli su
pokazali sposobnost značajnog približavanja stvarnim vrijednostima što bi u praksi moglo
značajno smanjiti rizik od prekomjerne ili nedovoljne nabave.This research focuses on analyzing beverage consumption at concerts, with a particular
focus on the impact of temperature conditions and musical genre on the consumption
of different types of beverages. By understanding these patterns, the research aims
to develop a tool that would allow catering companies to more accurately estimate the
quantities of beverages needed for events, with the aim of minimizing costs and optimizing
procurement.
Using machine learning models, such as Random Forest, XGBoost and neural networks,
the research evaluated their predictive ability based on historical consumption
data and augmented data. While perfect prediction is not possible, the models have
shown a significant ability to approximate actual values, which in practice could significantly
reduce the risk of over- or under-procurement
Four color theorem
No full textOvaj diplomski rad istražuje problem četiri boje, koji se odnosi na obojivost političkih karata tako da nijedne dvije susjedne zemlje ne dijele istu boju. Rad daje povijesni prikaz problema, objašnjava ključne računarske i matematičke dijelove dokaza Teorema o četiri boje te pokazuje utjecaj na razvoj teorije grafova i primjene znanstvenog računanja. Konačno, rad uključuje razvoj interaktivne web-aplikacije za bojanje karata i AI alat za personalizirane preporuke novih destinacija, što ilustrira praktičnu primjenu Teorema o četiri boje.This master thesis explores the Four Color Theorem, which addresses the problem of coloring of political maps such that no two adjacent countries share the same color. The thesis provides a historical overview of the problem, explains the key computational and mathematical components of the theorem’s proof, and demonstrates its impact on the development of graph theory and applications in scientific computing. Finally, the thesis includes the development of an interactive web application for map coloring and an AI tool for personalized recommendations of new destinations, illustrating the practical application of the Four Color Theorem
Anomaly tracking with panoptic models
No full textU ovom radu primijenjene su metode za praćenje anomalija. Za detekciju je korištena UNO metoda nadograđena na Mask2Former. Na temelju te detekcije isprobane su tri metode praćenja anomalija: SORT, SAM2 i SAM2MOT.In this work, anomaly tracking methods were applied. The UNO method added to Mask2Former was used for detection. Based on this detection, three anomaly tracking methods were tested: SORT, SAM2 and SAM2MOT
Web application for gyms
No full textRad opisuje razvoj i implementaciju web-aplikacije namijenjene teretanama. Aplikacija omogućuje korisnicima registraciju i prijavu (ako je dostupna). Treneri imaju mogućnost objavljivanja termina treninga, korisnici se mogu prijavljivati na njih, a administratori upravljaju registracijom trenera. Klijentski sloj aplikacije implementiran je pomoću biblioteke React, dok je serverski sloj razvijen uz korištenje Spring Boota.The paper describes the development and implementation of a web application intended for gyms. The application allows users to register and log in (if available). Trainers have the ability to publish training sessions, users can sign up for them, and administrators manage trainer registration. The client-side of the application is implemented using the React library, while the server-side is developed using Spring Boot
Computer vision system for the analysis of guitar playing
No full textU ovom radu razvijen je sustav za automatsku transkripciju gitarske tablature s fotografija koristeći metode računalnog vida. Sustav se sastoji od više povezanih komponenti:
segmentacije vrata gitare, geometrijskog procesiranja slike, detekcije žica i pragova, detekcije ruke, prstiju i njihovih vršaka, te konačno generiranja tablature. Evaluacija je pokazala da sustav postiže solidne rezultate u detekciji vrata, žica i pragova, kao i u prepoznavanju ruke i prstiju unatoč izazovima poput okluzija i varijacija u osvjetljenju. Glavni izazovi uključuju točnost detekcije tankih žica, probleme s razdvajanjem prstiju te nemogućnost transkripcije barre akorada. Ovaj rad uveo je poboljšanja u postojeće metode i pruža temelje za daljnja istraživanja i razvoj naprednijih modela.In this work, a system for automatic transcription of guitar tablature from photographs
using computer vision methods was developed. The system consists of several interconnected components: guitar neck segmentation, geometric image processing, detection of
strings and frets, detection of the hand, fingers, and fingertips, and finally tablature generation. Evaluation showed that the system achieves solid results in detecting the neck,
strings, and frets, as well as recognising the hand and fingers despite challenges such as
occlusions and lighting variations. The main challenges include accuracy in detecting
thin strings, issues with separating fingers, and the inability to transcribe barre chords.
This work introduced improvements to existing methods and provides a foundation for
further research and development of advanced models
Development and application of an algorithm for the detection and estimation of the position of 3D objects in the point cloud
No full textU radu je istraženo automatizirano uzorkovanje tla pomoću robotske ruke UR10e tvrtke Universal Robots koristeći sondu i dva različita svrdla. Za precizno upravljanje ruka je programirana u ROS-u s paketom MoveIt! za planiranje kretanja i Universal Robots ROS driverom. Posebni nosači i elektronički sklop omogućili su pričvršćivanje alata na robotsku ruku, uključujući DC motor za vrtnju svrdla upravljan relejem preko digitalnih izlaza ruke. Eksperimenti su provedeni na trima tipovima tla (mekana zemlja, prethodno obrađena i tvrda neobrađena zemlja) uz mjerenje sila i momenta na alatu. Svrdlo za drvo pokazalo se najboljim alatom za sve vrste tla, uz pouzdano prikupljanje uzoraka bez prekoračenja dopuštenih sila ili momenta. Sonda je često zapinjala u tvrdom tlu, a svrdlo za zemlju tražilo je velike momente bušenja i loše je zadržavalo zemlju zbog svog oblika, unatoč tome što je u nekoliko slučajeva prikupilo najveću masu uzorka. Zaključeno je da je svrdlo za drvo najučinkovitiji alat za automatizirano uzorkovanje tla u ovom sustavu, a da robotska ruka, iako vrlo fleksibilna, ne može generirati velike sile, poput specijaliziranih hidrauličkih ili električnih aktuatora, zbog čega nije optimalan izbor za sve tipove tla. Automatizacija ovakvog uzorkovanja tla može značajno smanjiti potrebu za ručnim radom i povećati učinkovitost terenskih analiza.The paper investigated automated soil sampling using a Universal Robots UR10e robotic arm using a probe and two different drills. For precise control, the arm was programmed in ROS with the MoveIt! motion planning package and the Universal Robots ROS driver. Special mounts and electronics enabled the attachment of the tool to the robotic arm, including a DC motor for rotating the drill controlled by a relay via the digital outputs of the arm. Experiments were conducted on three types of soil (soft soil, previously tilled and hard untreated soil) with measurements of forces and torques on the tool. The wood drill proved to be the best tool for all soil types, reliably collecting samples without exceeding the allowable forces or torques. The probe often got stuck in hard soil, and the earth drill required high drilling torques and had poor soil retention due to its shape, despite collecting the largest sample mass in several cases. It was concluded that the wood drill is the most efficient tool for automated soil sampling in this system, and that the robotic arm, although very flexible, cannot generate large forces, like specialized hydraulic or electric actuators, making it not an optimal choice for all soil types. Automation of this soil sampling can significantly reduce the need for manual labor and increase the efficiency of field analyses
An autonomous agent for playing a simple video game
No full textOvaj rad opisuje realizaciju autonomnog agenta koji igra jednostavnu videoigru, a treniran
je na principu podržanog učenja. U radu se najprije daje teorijski pregled podržanog učenja,
a zatim se detaljno opisuje programska implementacija kompletnog sustava; od
komunikacije s emulatorom koji pokreće igru pa do postupka treniranja i donošenja odluka.
U radu su također predstavljani rezultati treniranja agenta, komentirano je konačno
ponašanje agenta te su dani prijedlozi kako bi se sposobnosti agenta mogle poboljšati.This paper describes the implementation of an autonomous agent which plays a simple
video game, trained on the principle of reinforcement learning. The paper first provides a
theoretical overview of reinforcement learning and then describes in detail the software
implementation of the complete system; from communication with the emulator that runs
the game to the training and decision-making process. The paper also presents results of
the agent’s training, comments on the final behavior of the agent and provides suggestions
for improving the agent’s capabilities
Predicting player performance after transfers using machine learning and notational analysis
No full textOvaj rad bavi se razvojem modela strojnog učenja za predikciju statističkog učinka igrača nakon transfera u novi klub. Model koristi podatke o igračevim prethodnim performansama te karakteristikama momčadi, uključujući stil igre, kako bi predvidio njegove buduće statistike. Poseban naglasak stavlja se na primjenu notacijske analize u definiranju stilova igre timova, što omogućuje dublje razumijevanje konteksta u kojem igrač djeluje. Cilj rada je izgraditi prediktivni sustav koji može unaprijediti analizu transfera i podržati donošenje odluka u sportskim analitikama. Unatoč postojanju ograničenja poput relativno malih uzoraka i fokusa na kratkoročne trendove, rezultati pokazuju da je kvantitativno predviđanje performansi igrača nakon transfera moguće s umjerenim do dobrim stupnjem točnosti.This thesis focuses on developing a machine learning model to predict player performance statistics after transfers to new clubs. The model utilizes data on player's previous performances and team characteristics, including playing style, to forecast future statistics. Special emphasis is placed on applying notational analysis to define team playing styles, enabling a deeper understanding of the context in which the player operates. The aim is to build a predictive system that can enhance transfer analysis and support decision-making in sports analytics. The thesis includes program source code, results with necessary explanations, and relevant literature. Despite the existence of limitations such as relatively small samples and focus on short-term trends, the results show that quantitative prediction of player performance after transfers is possible with moderate to good levels of accuracy
Patially Observable Markov Decision Processs and Novelty-Biased Consensus in Multi-Agent Systems for Active Perception
No full textOvo istraživanje bavi se primjenom parcijalno osmotrivih markovljevih procesa od-
lučivanja (POMDP) i konsenzusa pristranog najnovijoj informaciji (novelty-biased con-
sensus) u višeagentskim sustavima za aktivnu percepciju. Razvijen je decentralizirani
pristup u kojem svaki agent lokalno rješava svoj POMDP model, dok se razmjena infor-
macija među agentima ostvaruje putem konsenzus protokola. Poseban naglasak stavljen
je na analizu utjecaja parametara kvalitete senzora i težina u konsenzus algoritmu na
učinkovitost kolektivnog zaključivanja. Provedeni su simulacijski eksperimenti na stan-
dardnim problemskim scenarijima, gdje su varirani parametri pristranosti i točnosti sen-
zora. Rezultati pokazuju da NB konsenzus omogućuje bržu razmjenu informacija i brže
donošenje odluka, osobito u uvjetima niske pouzdanosti senzora, dok je sustav najrobus-
tniji kada je modelirana točnost senzora usklađena sa stvarnom. Analizom node-based
i edge-based pristupa utvrđeno je da node-based pristup osigurava veću stabilnost, dok
edge-based može ubrzati širenje informacija, ali uz rizik od oscilacija. Predloženi okvir
pokazuje potencijal za primjenu u naprednim autonomnim sustavima i otvara moguć-
nosti za daljnja istraživanja.This research deals with the application of partial eight-way Markov decision pro-
cesses (POMDP) and novelty-biased consensus in multi-agent systems for active percep-
tion. A decentralized approach was developed in which each agent solves its POMDP
model locally, while the exchange of information between agents is realized through a
consensus protocol. Special emphasis is placed on the analysis of the influence of sensor
quality parameters and weights in the consensus algorithm on the efficiency of collec-
tive reasoning. Simulation experiments were carried out on standard problem scenarios,
where the bias and accuracy parameters of the sensors were varied. The results show
that NB consensus enables faster information exchange and faster decision-making, es-
pecially in conditions of low sensor reliability, while the system is most robust when
the modelled sensor accuracy is aligned with the real one. By analysing node-based and
edge-based approaches, it was determined that the node-based approach provides greater
stability, while the edge-based approach can accelerate the spread of information, but
with the risk of oscillations. The proposed framework shows potential for application in
advanced autonomous systems and opens up opportunities for further research