Chalmers Open Digital Repository

Välkommen till Chalmers öppna digitala arkiv!

Här hittar du:

Studentarbeten utgivna på lärosätet, såväl kandidatarbeten som examensarbeten på grund- och masternivå
Digitala specialsamlingar, som t ex Chalmers modellkammare
Utvalda projektrapporter

Forskningspublikationer, rapporter och avhandlingar hittar du i research.chalmers.se

Enheter i Chalmers ODR

Välj en enhet för att se alla samlingar.

Visar 1 - 2 av 2

Senast publicerade

Structuring Educational Media into a Searchable Digital Database with API Access - A prototype for educational media management
(2026) Ramsten, Hampus; Tuovinen, Joakim
Digimar is an educational platform for maritime communication that uses several types of learning material, such as videos, transcripts, presentations, documents, images and chatbot content. Before this project, the material lacked a centralized technical interface, which made it harder to search, organize and reuse content across different frontend applications. This thesis presents the design and implementation of a searchable digital database with API access for Digimar. The system consists of a Go based HTTP REST API, a PostgreSQL database, a static frontend console and a local Docker Compose setup. The API supports several resource types, including chapters, sections, content items, videos, transcripts, PowerPoint files, documents, images and chatbots. It also includes search across metadata and text fields, role based API keys, health and readiness endpoints, and documentation for the API and data model. The result is a working prototype that provides one shared backend for managing Digimar learning material. Most of the planned goals were completed, including storage of the planned content types, REST API access, search, access control, Docker based local development, documentation and a frontend tool for manual operation. The planned PDF to markdown conversion automation was not completed and remains future work. Other future improvements include more formal database migrations, improved file storage, search ranking based on user preferences and production concerns such as backups, caching and monitoring.
Manifest-Level Permission Debloating on Android Applications
(2026) De Pinto, Sotiri
Modern Android applications frequently contain functionality that most users never need or use, a phenomenon known as software bloat. One specific manifestation of bloat is permission over-declaration, where applications declare permissions in their manifest whose corresponding Android API methods are never invoked in the application’s code. Over-declared permissions mislead users, unnecessarily expand the attack surface of the application, and may signal access to sensitive resources such as the camera or microphone without any functional justification. This thesis proposes and evaluates a two-phase hybrid debloating pipeline for Android applications. The first phase is a custom static analysis tool that automatically detects over-declared permissions by scanning the application’s smali bytecode for permission-protected API calls, and removes unused permission declarations from the manifest without modifying any executable code. The second phase applies MiniMon, an existing monitor-based debloating framework, to the output of the first phase, removing methods that were not exercised during monitored usage sessions. The pipeline is evaluated on a dataset of 38 real-world Android applications. The results show that permission over-declaration is widespread, affecting 44.7% of the applications in the dataset. The first phase reduced APK size in 16 out of 17 affected applications with an average reduction of 122.14 KB, while preserving functional correctness by construction. The combined pipeline produced additional size reduction over monitor-based debloating alone in 11 out of 28 applications, demonstrating that the two approaches are complementary: permission-based debloating addresses manifest-level bloat that is invisible to usage-based approaches, while monitor-based debloating addresses code-level bloat that manifest analysis cannot reach.
Automated guided vehicle localization: A mapping strategy in dynamic environments, using computer vision
(2026) Cederberg, Albin; Wallden, Viktor
Many warehouses rely on automated guided vehicles (AGVs) that use 2D-LiDAR based localization and static environmental landmarks. In dynamic warehouse environments, this approach can become unreliable due to occlusions and changes of previously considered static landmarks. A vision based localization method is able to detect features in the environment that a 2D-LiDAR cannot. Therefore, this thesis aims to answer the main question: "Can a visual localization method improve localization accuracy compared to LiDAR-based 2D-localization, based on common accuracy metrics like absolute trajectory error (ATE)?" To evaluate our own solution, a baseline was created consisting of a modified version of the open-source method pySLAM [1]. The baseline was modified to rely on a static map for localization, reflecting how many industrial navigation systems operate. Based on the observed limitations of this approach in dynamic environments, a new method called 2P (two point)-SLAM was developed and made available as open source [2]. The method separates map points into static and mutable points, allowing the system to preserve a pre-recorded static map while adapting to environmental changes by adding and removing mutable points when needed. The results show that the maximum error of the baseline is 307 mm during the dynamic test. SLAM achieves a maximum error of 144 mm, while 2P-SLAM achieves 184 mm. However, 2P-SLAM adds 83% fewer points than SLAM. The ATERMSE suggests that there is no significant difference between 2P-SLAM and the baseline, with values of 76 mm and 78 mm respectively. As expected, the existing 2D-LiDAR solution struggles in the dynamic environment with an ATERMSE of 309 mm. The conducted experiments therefore suggest that a vision based localization method can improve localization accuracy compared to the evaluated 2D-LiDAR localization method in the tested dynamic environment. Since each localization method was evaluated using a single execution and the visual localization pipeline exhibits stochastic behaviour, the reported results should be interpreted as indicative rather than statistically significant. Further testing and repeated evaluations are required to strengthen these findings.
Learning-Based Task Assignment for Automated Guided Vehicles: Applying graph neural networks to optimise task assignment in an online warehouse environment
(2026) Borgvall, Axel; Ivarsson, Nils
Efficient task assignment in multi-Automated Guided Vehicle (AGV) warehouse environments is critical for optimizing industrial logistics. In collaboration with MAXAGV, this thesis evaluates the application of reinforcement learning (RL) to address this challenge. The warehouse environment is modelled as a graph, and a Graph Neural Network (GNN) policy is trained using Proximal Policy Optimization (PPO) to assign tasks to the vehicle fleet. To capture the complex topology of the facility, which is characterized by long-range spatial configurations and lock-relations that limit standard embedding methods like Node2Vec, a novel transductive node embedding scheme trained via multiple task-specific decoders is introduced. Three core GNN architectures: Graph Convolutional Networks (GCN), Graph Attention Networks (GAT), and Graph Transformers, along with their heterogeneous extensions, are evaluated and compared against conventional heuristic baselines. The empirical results demonstrate the performance trade-offs between the learning-based architectures and traditional heuristics. Furthermore, the study addresses the broader challenges of deployment, specifically the complexities of reward shaping in real-world logistics systems and the systemic barriers to integrating learning-based methods into legacy industrial infrastructures.
Feature Extraction and Classification of Knee Motions Using a Sensor-Equipped Orthosis
(2026) Bengtsson, Alicia; Bursell Palm, Hugo
This thesis investigates the feasibility of using a sensorized knee orthosis for movement classification and sensor-derived feature extraction during controlled lowerlimb movements. The orthosis was equipped with two six-axis IMU sensors and an absolute capacitive rotary encoder. A movement data-collection study involving 24 healthy participants was conducted, in which four lower-limb movements and three intentionally simulated execution variations were recorded. The collected data were used to train and evaluate three neural network architectures commonly applied to time-series classification: CNN-LSTM, LSTM-CNN, and a parallel architecture. The study is limited to controlled motion analysis using a sensorized wearable system and does not evaluate medical, diagnostic, therapeutic, rehabilitation, or clinical outcomes. The results show that the sensor system provided stable sampling frequencies and data suitable for classification of four movements. The CNN-LSTM model achieved a mean leave-one-subject-out (LOSO) classification accuracy of 99.42 ± 0.35% for the predefined movement classes and more than 99% pseudo-live classification accuracy after post-processing. In addition, movement-related features such as range of motion, knee angle deviations, movement smoothness, and repetition count were extracted from the recorded sensor data. A modified CNN-LSTM architecture was also evaluated for classification of predefined execution categories, demonstrating the feasibility of distinguishing between intentionally defined movement variations. The feature-extraction results should be regarded as exploratory because no external kinematic ground truth was available. Similarly, the execution-category classification represents predefined movement categories rather than a validated assessment of movement quality. Overall, the findings demonstrate the potential of wearable sensor systems and machine learning for controlled movement classification and sensorderived movement analysis.