Civilingenjör i datateknik på Utexpo

A Bloat-Free 3D Game Engine

• Participants: Edvin Andersson and Gustaf Andersson.
  

The complexity and performance impact of software has significantly increased over time. Between 1976 and 2002, microprocessor performance grew by 52 percent annually (Moore’s law), allowing the software industry to overlook design patterns and performance aspects. However, as Moore’s law slows, it can be argued that software performance is becoming more prevalent and should be considered more of a software problem instead of a hardware problem.

To address this, our project aimed to develop and evaluate a bloat-free 3D game engine, prioritizing memory efficiency, loading times, and overall code efficiency. Using C++ with specific libraries, language features, and design patterns, we applied these techniques to create a game engine. Game engines simplify and speed up game development by providing essential systems like rendering, physics, events, and input management.

We identified four key problem areas: memory utilization, CPU usage, export time, and engine launch time, guiding the development of our game engine, CogWheel. Our testing strategy included comparing CogWheel's performance against industry standards like Unity and Unreal Engine. Results for CogWheel were promising, showing significant improvements in export times, CPU usage, and memory utilization, particularly in simpler projects. However, performance differences in complex scenes highlighted areas for further development and testing.

Auto Generate CAD Drawings From Text Descriptions

• Deltagare: Abd Alrahman Atieh.
  

Projektet fokuserar på att utveckla ett verktyg för att skapa 3D-modeller genom att använda en stor språkmodell (LLM) och Blender Python API. Målet är att göra processen med att skapa 3D-modeller tillgänglig för människor med olika erfarenhetsnivåer. Istället för att kräva djupgående kunskap om 3D-modelleringstekniker kan användarna beskriva önskade modeller i textform, och sedan genereras motsvarande kod automatiskt. För att utvärdera verktyget jämförs den genererade koden med kod genererad av andra språkmodeller. Slutligen visar resultaten att verktyget kan skapa olika 3D-modeller och scener effektivt, vilket gör det till en användbar resurs för dem som behöver snabbt skapa 3D-innehåll utan avancerad expertis. Valet av codeLLaMA motiveras av dess kostnadsfrihet, öppen källkod och möjlighet att användas offline.

Blender Interface Integration with a Website

• Deltagare: Abdulfattah Morad och Ali Abedrabbo.
  

Vårt projekt fokuserar på att utveckla en användarvänlig webbapplikation för enkel rendering och interaktion med 3D-modeller. Genom att integrera olika teknologier, inklusive JavaScript, Blender och Docker, strävar vi efter att skapa en flexibel plattform där användarna kan utforska och skapa i 3D-miljöer. Med ett användarcentrerat tillvägagångssätt siktar vi på att leverera en lösning som möjliggör smidig skapande och utforskande av 3D-världen.

Early Detection and Management of Hearing Loss

• Participants: Neamah Jameel and August Melén.
  

This thesis focuses on the development and evaluation of a portable device based on Arduino technology for the early detection and management of hearing loss. The device is programmed using IAR Embedded Workbench, an environment suited for microcontroller applications, to simulate the assessments typically performed at hearing clinics. The primary goal is to make hearing tests more accessible, particularly in workplace settings, allowing individuals to conduct preliminary evaluations of their hearing efficiently and effectively.

The device measures hearing sensitivity across a range of frequencies and can identify common patterns of hearing loss and present these findings in a user-friendly interface. The results are validated through comparisons with traditional clinical tests and further analyses based on data from various age groups.

Key aspects of the project include the technical design and implementation of the hearing test, the rationale behind the choice of materials and technologies, and the calibration techniques used to ensure accuracy. The device's performance, while promising, highlights the need for further validation with larger sample sizes to enhance its accuracy and reliability.

Overall, this project contributes to the field of auditory health by providing a tool that facilitates early detection of hearing loss, aiming to increase regular auditory health monitoring and raise awareness about hearing conservation.

Enabling Communication: Instantaneous Translation from Sign Language to Text

• Participants: Theo Johnsson and Isak Karlsson.
  

This study explores Convolutional Neural Networks (CNNs) in detail, this includes various layers and architectural designs. It highlights the creation of a dataset for the Swedish Sign Language (SSL) and the use of augmentation techniques to enhance model training. The dataset consisted of 47 320 images. The project uses hand-tracking to locate the sign for translation. Furthermore, the models included a fine-tuned MobileNet model and a custom model. Notably, fine-tuning MobileNet's architecture achieved the highest test accuracy of 96 %. Additionally, the research evaluates the applicability of image recognition models on low-power devices, exemplified by a Raspberry Pi 4 model B for practical experimentation. Through these processes insights into the efficacy of CNNs and their potential deployment on low-power platforms are analyzed.

Exploring Particle Dynamics: A Comparative Analysis of Traditional Simulations and Neural Network Predictions

• Participant: Joppe Doris Daniël Smink-Katagiri.
  
• Collaboratior: Avalanche Studios Group.

Recent research within granular computing has made it possible to simulate real-world scenarios using particles.

While particle systems were originally developed to model fuzzy objects in computer graphics, such as clouds, flames or raindrops, the abstract idea of points moving through space can be applied to nearly anything. This includes research and scientific advancements in, among others, the following areas:

• Fluid flow and turbulence simulations.
• Ecosystem simulations, such as colony stability and resource sharing.
• Urban planning and congestion prevention by simulating traffic flow.
• Self-driving cars.
• Planetary movement predictions.
• Medical development and drug design.

However, while the theories using particles seem promising, the applications may become limited on a larger scale. Simulations with a high particle count, especially those with inter-particle collisions, can be computationally expensive.

One solution is Neural Networks. There have been cases where Neural Networks achieved results up to 100 000 times faster than traditional numerical methods. Furthermore, because of the so-called Universal Approximation Theorem, a multitude of networks have been proven to be able to approximate nearly any function.

This work explores two networks: the Vanilla Network and the custom designed Modular Network. While the results produced are not identical to the source data, the conclusions of this project have shed light on the field, providing potential subsequent steps.

Here I go: A prediction model for e-bike and e-scooter positioning inside a CCAM environment

• Participants: Ruben Croall and Douglas Jonsson Lundqvist.
  

This thesis presents a prediction model for e-bikes and e-scooters, aimed at enhancing traffic safety and efficiency by sharing their intentions of future possible positions among road users. The research addresses the current automated vehicle technologies which lack communication with road users. The prediction model is based on and tested with a mobility model, adapted for modelling e-bikes and e-scooters in a simulator program primarily used for pedestrians. This implementation has produced the ability to predict future positions and further the development of intention-sharing capabilities in urban traffic scenarios. The model is built upon physical parameters and mathematical models for a controlled and regulated model. Polynomial regression was applied to predict positions based on historical data and the results were evaluated with RMSE metrics, demonstrating the prediction accuracy in different scenarios.

The thesis also includes the integration of the prediction model into a hardware setup, a Raspberry Pi. Demonstrating the practical application and retaining the effectiveness of the model in a real-time environment. With this, the research highlights the possibility of implementing this in CCAM systems. The results show promising accuracy with a simple controlled model using as little necessary data as possible. The project work contributes to the field of intelligent transport systems by providing a scalable solution to enhance the interaction between VRUs and vehicles, getting a step closer to achieving the Vision-Zero goal of having zero traffic related accidents or fatalities.

Improving audio intelligibility in intercom devices

• Deltagare: Hieu Tran och Thomas Lundqvist.
  
• Samarbetspartner: Axis Communications.

Porttelefoner används ofta i högljudda miljöer. Ett exempel på en sådan miljö är vindutsatta områden, där operatören i ett rum kan uppleva svårigheter att tydligt uppfatta tal från användaren som talar i en porttelefon på grund av den omgivande höga ljudnivån.

Detta projekt utfördes vid Axis Communications, ett svenskt företag inom nätverksbaserade lösningar för videoövervakning och fysisk säkerhet. Projektets mål var att utforska och implementera ett adaptivt filter med en adaptiv algoritm i C-programmering för att förbättra ljudkvaliteten genom att minska bruset hos porttelefoner i utmanande miljöer. Genom att använda ett lämpligt adaptivt filter strävade projektet efter att minska brus och optimera ljudkvaliteten.

Efter noggranna studier valdes NLMS-algoritmen för implementering i projektet. Algoritmens prestanda utvärderades genom att analysera beräkningstid, medelkvadratfelet (MSE – Mean Squared Error) och signal-brus-förhållandet (SNR - Signal-to-Noise Ratio) i Matlab, samt genom användartester för att säkerställa kvaliteten.

Detta projekt uppnådde målen genom att utveckla ett fungerande adaptivt filter. Under projektets gång hanterade systemet kontinuerligt dataströmmar effektivt i praktiska tester, vilket bekräftade att det fungerade utan fördröjningar. Detta bevisade det adaptiva filtrets effektivitet i verkliga applikationer, särskilt i högljudda miljöer.

Industrial Wi-Fi Redundancy Methods

• Participants: Oliver Dizdarevic and Fabian Henrysson.
  
• Collaborator: HMS Anybus.

This thesis explores various Wi-Fi redundancy methods to enhance network resilience in industrial settings. Maintaining uninterrupted data transmission is crucial due to the increasing reliance on wireless technologies for industrial operations. Our research investigates the performance of different redundancy strategies, including Multi-Link Operation (MLO) and Truncated Automatic Repeat Request (TARQ), through practical prototyping and testing on specific hardware configurations. We assessed these methods’ effectiveness in mitigating packet loss and improving transmission consistency under varying attenuation.

Network-assisted positioning in confined spaces

• Participants: Emelie Leifsdotter and Franjo Jelica.
  

This project compared and evaluated the accuracy of two signal strength-based trilateration methods using either Wi-Fi or Bluetooth Low Energy (BLE). The purpose of evaluation is part of the long-term vision of improving the safety of workers in confined environments, such as factories, where positioning systems like GPS are not suitable due to signal obstruction. Samples of signal strengths to three reference nodes were collected at randomized positions within a confined space, using off-the-shelf hardware. It was observed that signal strength using Wi-Fi achieved a better accuracy of predicting the actual position within the testing environment than signal-strength using BLE. However, because of the power efficiency of BLE it is a viable candidate for a future low-cost and device-based Indoor Localization System to potentially be used and worn by workers. The results while aligned with similar existing literature, infer what a low-cost indoor positioning system might achieve.

This project is confidential and is placed in room I105 at Utexpo.

Sustainability Data Extraction System

• Participants: Dalia Abbas and Nashwa Nana.
  
• Collaborator: ICONSOF.

Sustainability reporting is essential for organizations to report their environmental, social, and governance (ESG) performance. However, extracting and structuring data from sustainability reports can be challenging, leading to inefficiencies and inconsistencies. This project aims to develop an integrated system for sustainability reporting by leveraging artificial intelligence (AI) techniques, particularly natural language processing (NLP), to extract and structure data from sustainability reports. Utilizing the GPT-3 model by OpenAI, the system processes unstructured text from PDF reports into a structured format compliant with European Sustainability Reporting Standards (ESRS). Through a meticulously designed pipeline, including PDF parsing, text sanitizing, batching, parallel API requests and postprocessing, the system achieved efficient and accurate extraction of key sustainability information. The system's effectiveness is evaluated using cosine similarity metrics, comparing model outputs with manually extracted data. The results demonstrate high alignment between the model outputs and manual extractions, validating the system's performance. This project contributes to advancing sustainability reporting practices, providing organizations with a robust tool for transparent and standardized disclosure of environmental, social, and governance (ESG) impacts.

The effect of a mixed-capability vehicular fleet on Vulnerable Road User safety

• Participants: Nicholas Sjögren and Duc Huy Vu.
  

This thesis investigates the integration of vehicles with differing levels of automation and connectivity within suburban traffic systems, focusing on their impact on road safety and traffic efficiency. Employing a Cooperative and Connected Automotive Mobility (CCAM) framework, the study examines how vehicles that share real-time information and intentions under various CCAM configurations influence the dynamics of suburban mobility. Utilizing simulation tools such as SUMO and Artery, this research conducts multiple scenario simulations to capture the interactions between automated, connected, and conventional vehicles. The simulations target the implementation of Intelligent Transport Systems (ITS) protocols (ETSI ITS-G5), assessing their effectiveness in enhancing safety and efficiency in suburban environments. The analysis quantifies the direct impacts of vehicle automation on performance metrics such as collision frequency and traffic flow. The findings provide insights into deploying advanced vehicular technologies, ensuring their beneficial integration into increasingly complex suburban traffic networks, thus supporting global road safety initiatives like Vision Zero. This research contributes to the ongoing discourse on vehicular technology by offering evidence-based recommendations for optimizing the integration of mixed-capability vehicles in public roadways, prioritizing safety and efficiency to accommodate all users in the evolving transportation landscape.