Eye tracking is a cornerstone technology for next-generation human-computer interaction, particularly in Extended Reality (XR), and other healthcare applications. However, traditional frame-based eye tracking systems are constrained by latency, power consumption, and motion blur. Event cameras offer a promising alternative with their high temporal resolution, high dynamic range and low data redundancy, but existing event-based methods often struggle to balance tracking accuracy, computational efficiency, and robustness, especially on resource-constrained mobile hardware.

This thesis addresses these challenges by proposing a novel, purely event-based eye tracking pipeline designed for high-frequency performance and robust accuracy within a strict computational budget. The pipeline accepts only event streams and estimates the pupil region in the field of view. The core contribution is a dual-state framework that synergistically combines a deep learning-based pupil detector with a lightweight, rapid template updater. For robust detection, a lightweight, attention-augmented segmentation network, named PupilUNet, is developed. It leverages a truncated MobileNetV3 Small encoder and a parameter-free attention mechanism to accurately segment the pupil boundary from Speed-Invariant Time Surface (SITS) representations, which provide a stable input by normalizing for motion speed. To overcome the scarcity of annotated data, a comprehensive framework is introduced to augment a large-scale training dataset from limited initial labels. Once a high-confidence pupil template is detected, the system transitions to a rapid updating mode, employing an optimized, vectorized point-to-edge matching algorithm to track the pupil at
kilo-Hertz frequencies with millisecond latency. A dynamic control logic monitors tracking quality and seamlessly reverts to the robust detection mode when necessary, ensuring both speed and resilience.

Experimental results on the EV-Eye dataset validate the pipeline’s effectiveness. The PupilUNet detector achieves a P5 accuracy of 96.3% (pupil center error < 5 pixels), while the rapid updater operates with an average latency of approximately 1 ms. The lightweight PupilUNet model contains merely 0.177 M parameters and inferences within 0.553 GFLOPs. The fully integrated system sustains a P5 accuracy of 85.2% while achieving a peak tracking frequency of over 960 Hz. This work demonstrates a practical and efficient solution that successfully navigates the trade-offs between accuracy and latency, establishing a new baseline for high-performance, event-based eye tracking on mobile and embedded systems.

With the increasing demand for artificial intelligence (AI), intelligent systems have become deeply integrated into various aspects of modern life, including autonomous driving, smart assistants on mobile devices, and powerful online language models such as ChatGPT. In addition, the emergence of generative models for text and image synthesis has significantly reduced the cost of ac cessing and interacting with information. However, the advancement of these AI applications comes at the cost of ever-growing computational and memory requirements. These requirements pose substantial challenges even for high end computing systems, and become prohibitive when deploying AI models on resource-constrained platforms such as embedded devices and Internet of Things (IoT) nodes. This thesis presents a distributed inference framework for Transformer mod els where we design a fine-grained, channel-wise parameter partitioning scheme. Importantly, the implementation of our framework is also independent of con ventional AI frameworks such as PyTorch, making it efficient, portable, and adaptable to virtually any compute-capable device. We begin by analyzing the computational and memory limitations of main stream hardware platforms, highlighting the motivation to aggregate multiple low-power devices to collectively execute AI workloads. Through software based simulation, we validate the correctness of the partitioned inference scheme and demonstrate that it introduces no functional deviation from unpartitioned single-device execution. The simulation also facilitates precise estimation of data flow and compute demand across multiple collaborating devices. Furthermore, we introduce a full-stack load balancing algorithm that enables adaptive task allocation based on heterogeneous hardware specifications, taking into account factors such as bandwidth, memory capacity, and communication latency. In summary, this thesis proposes a split, practical, and high-granularity Trans former inference framework that is compatible with heterogeneous hardware configurations, offering a promising step toward enabling AI inference distributively on a network of resource-constrained embedded platforms.

This thesis researches mitigations for BLE relay attacks. A design for a timebased distance bounding protocol using the Bluetooth channel sounding feature introduced in the new Bluetooth 6.0 core specification is presented. Bluetooth channel sounding is compromised of two distance measurement techniques: Phase-Based Ranging (PBR) and Round Trip Tim (RTT). The proposed protocol requires consistent channel sounding distance measurements in order to minimize the likelihood of succesfull relay attacks. Single-antenna channel sounding measurements have shown poor spatial and sequential consistency in a complex multipath office environment. In order to overcome inaccuracies that arise due to multipath propagation, this thesis investigates the optimal antenna configuration for Bluetooth channel sounding using multiple antennas. A comparison
between the root-mean-square error and maximum error of the single-antenna baseline and the proposed multi-antenna solution for both spatial and sequential consistency in a complex multipath office environment shows that there is, on average, a 58% reduction in error metrics when the optimal multi-antenna setup is used. The performance of the optimal multi-antenna channel sounding setup
in the complex environment approaches the single-antenna baseline performance
in an ideal outdoor environment. This shows that the added antenna diversity
successfully overcomes the negative effects due to multipath propagation.

This research explores the feasibility of implementing lane detection on lightweight microcontrollers using a combination of traditional image processing and compact machine learning methods. With the aim of enabling real-time inference under strict hardware constraints, several models were trained and evaluated against a custom image processing pipeline. Each approach was tested for accuracy, speed, and resource usage on the Raspberry Pi Pico 0 microcontroller. While these solutions fall short of cutting-edge accuracy and cannot process as much information as state of the art models, their low cost, minimal power consumption, and real-time performance highlight their potential. These findings suggest that lightweight lane detection is a viable direction for further research in embedded autonomous systems.

Visible light positioning (VLP) systems are a promising solution for indoor positioning, utilizing light-emitting diodes (LEDs) as transmitters and photodiodes (PDs) as receivers.
A received signal strength (RSS) based VLP system's accuracy is heavily dependent on the density of collected fingerprints, being a very labor-intensive process.
In this study, we focus on RSS fingerprints to achieve centimetre level positioning accuracy, while addressing the challenges of labor-intensive fingerprint collection and deployment on resource-constrained devices like the Raspberry Pi Pico microcontroller.
We found different neural network architectures using Neural Architecture Search (NAS) to optimize the VLP system, which achieve on average $12mm$ positioning error with low inference latency around $50ms$ on the Raspberry Pi Pico.

Real-time traffic sign recognition on microcontrollers introduces challenges due to limited memory and processing capacity. This study investigates the trade-offs between model size, classification accuracy, and inference latency within hardware constraints. We present an efficient network architecture called AykoNet with two variants: AykoNet-Lite, prioritizing model size and inference latency, and AykoNet-Pro, prioritizing classification accuracy. We trained AykoNet on the German Traffic Sign Recognition Benchmark (GTSRB) and specifically optimized it for deployment on the Raspberry Pi Pico microcontroller. AykoNet-Lite delivers 94.60% accuracy with only a 36.80KB model size and 55.34ms inference time, while AykoNet-Pro achieves 95.90% accuracy with an 80.18KB model size and 87.13ms inference time. Our approach demonstrates the effectiveness of domain-specific preprocessing and architectural design, class-aware data augmentation, and the strategic use of depthwise separable convolutions. These results validate the feasibility of real-time traffic sign recognition in resource-constrained embedded systems. Specifically, AykoNet-Lite strikes an optimal balance between model size, classification accuracy, and inference latency for practical deployment in autonomous navigation applications.

Visible light positioning (VLP) enables accurate indoor localization by leveraging a dense deployment of LEDs in future lighting infrastructure, but its widespread adoption is hindered by two key challenges: the need for densely sampled fingerprint datasets and performance degradation due to LED aging or failure. In this work, we propose a VLP framework that reduces reliance on dense fingerprinting and remains robust over time without requiring manual re-fingerprinting. Using a dataset acquired from the DenseVLC testbed, we evaluate preprocessing techniques that enhance positioning accuracy under noisy received signal strength (RSS) measurements. To address long-term reliability, we introduce a simulation framework that models LED degradation and sudden failures. Most importantly, we present an online learning approach that dynamically adapts the positioning model in response to environmental and infrastructure changes.
In our simulations, this approach maintains the original level of accuracy despite aging effects. In some cases, it yields up to a 95% improvement when evaluated over longer timespans. Furthermore, our preprocessing contributions have led to a 30% improvement to baseline performance without aging. Our results demonstrate a path toward scalable, self-sustaining VLP systems suitable for real-world deployment.

This work investigates the feasibility of performing monocular depth estimation on highly resource-constrained hardware, specifically the Raspberry Pi Pico Zero microcontroller. In contrast to existing approaches that rely on large convolutional networks and high performance devices, this study explores a set of custom lightweight encoder-decoder architectures, including one inspired by L-ENet, L-EfficientUNet, μPyD-Net, and an LSTM-μPyD-Net combination, designed to operate within strict memory limits. These models were trained on a preprocessed KITTI dataset, with either LiDAR depth maps or SGM (Semi-Global Matching) dense depth maps, and evaluated in terms of accuracy, model size, and real-time inference performance. Results demonstrate that meaningful depth prediction is achievable on microcontrollers, paving the way for low-cost autonomous navigation systems and broader applications of TinyML in embedded robotics, with SGM proving to be the best preprocessing technique, and the LSTM-μPyD-Net having the best accuracy when trained on the full Train split of the KITTI dataset.

The advancement in transparent screen technology has promoted adoption of full-screen design on mobile devices, reducing the area occupied by optical sensors to maximize the devices' screen-to-body ratio. In modern smartphones, front-facing optical sensors, such as ambient light sensor and camera, now must be placed under the transparent screen to capture ambient light and visual information. Motivated by this trend, we propose Through-Screen Computing in this dissertation. It is a new concept that refers to the processing of light signals for various computing purposes such as communication, sensing, and imaging, where the light comes from the physical world and passes through a special medium -- the transparent screen -- before reaching the under-screen optical sensors. This concept opens up new challenges and opportunities in connectivity, privacy, and security of future devices equipped with transparent screens. In this dissertation, we outline a vision for through-screen computing and address the challenges of transparent screens acting as both passive blockers and active interferers of input light signals.

This dissertation focuses on two subsystems in the context of through-screen computing: Through-Screen Visible Light Communication (VLC) and Screen Perturbation for Visual Privacy Protection. In the context of VLC, the full-screen trend challenges the deployment of this technology. We propose Through-Screen VLC with under-screen optical sensors as receivers. To address the attenuation of the light by the transparent screen, we develop SpiderWeb, a system exploiting the color domain to mitigate the color-pulling effect introduced by the transparent screen. We also leverage the Under-Screen Camera (USC) for VLC and design novel demodulation algorithms to reduce multi-pixel screen interference and improve performance. Experimental results show significant improvements in both data rate and transmission range, using  a prototype we build with two commercial smartphones. For privacy protection, we propose Screen Perturbations, modifying pixels displayed on the transparent screen to embed speckled color blocks and color shifts in the final image captured by the USC. Screen perturbations can be exploited to disrupt advanced deep neural networks used on image classification and face recognition tasks. We first design two image-specific methods to add screen perturbations to the images captured by USC. Next, we develop Unicorn, a universal screen perturbation method optimized for robustness in various conditions. All these designed perturbations work successfully against various deep neural network-based image classification services with high success rates.

Through these two subsystems, as well as the proposed theoretical and experimental approaches and results, we demonstrate the transformative potentials of through-screen computing, setting the stage for future research and development on various computing purposes in the era of transparent screen and full-screen devices.

Federated learning (FL) allows multiple clients to train a machine learning model on a server without sharing their private data. To reach a consensus, the server collects alternative information such as model updates. The sub-field of heterogeneous FL investigates scenarios where clients have varying charac-teristics. This thesis investigates the model-heterogeneity challenge, where the clients train different model architectures, leading to compatibility issues in reaching a consensus. This thesis addresses model heterogeneity in a setting where each client owns multiple devices. We also assume that devices belonging to the same client can share data among themselves. This assumption is atypical in FL and prompts us to re-assess the limitations of knowledge transfer techniques such as knowledge distillation. We propose a set of tools centered around knowledge distillation that leverage the data-sharing policy to address model heterogeneity. We also quantify the levels of data and device heterogeneity present on the devices to
develop an adaptive solution that fits the type of heterogeneity encountered. We demonstrate the effectiveness of our solution in training a machine learning model in a federated setting with model heterogeneity by comparing it with alternative solutions in terms of latency and accuracy. Finally, we provide dir-ections for further research, such as devising tests for data heterogeneity and exploring other model compression techniques in heterogeneous FL.

Unmanned aerial vehicles (UAVs), often referred to as autonomous drones, are becoming more and more prevalent in our daily lives. Drones are usually equipped with traditional frame-based cameras and have functions like object detection. However, the high energy consumption of frame-based cameras presents a challenge to drone endurance. In addition, due to its own hardware limitations, problems such as blurring and deformation will occur when capturing high-speed moving objects. In contrast, event cameras, as one of the latest neural technology cameras, have the characteristics of low power consumption, low latency, and sensitivity to high-speed objects. This makes them well-suited for integration into drone platforms. In this thesis, we introduce a method for high-speed moving object recognition based on event cameras. This approach involves enhancing objects with externally added ``common features": fiducial markers, and employing a custom-developed deep learning neural network to detect these markers. These marks can carry some messages, such as the information of related objects. When these marks are detected, it is considered that objects are detected. We also developed a fiducial marker decoding method based on region segmentation to obtain the message content, thereby achieving interaction with the detected object. Evaluation results show that the proposed method has a low computing time of 26.9 ms, low storage and memory usages of 670 MB CPU memory and 750 MB GPU video memory, and a high accuracy of 77.9%, making it suitable for high-speed object recognition based on event cameras.

The increasing expansion of the electrical network accentuates the need for a better understanding of the quality of the existing infrastructure. Assessing the quality of the individual cables becomes instrumental in prioritizing replacements and grid reinforcements within the network. A critical evaluation of insulation quality in paper insulated lead covered cables is important to ensure a robust electricity grid. Moisture plays a key role in the degradation of the paper insulation. In this thesis a novel microcontroller based near-infrared spectroscopy system design is proposed which is able to measure the moisture content of paper insulated lead covered cable insulation with a theoretical accuracy of about ±1% moisture content, while being cost effective and portable. It achieves this accuracy by making use of an outlier detection algorithm which is able to increase the accuracy of the system from ±3% to ±1%, by removing outliers of individual sensors and averaging the results of multiple measurements. The resulting output of the algorithm is linked to the asset number of the paper insulated lead covered cable under test, and uploaded to the cloud to enable further data analysis. The proposed near-infrared spectroscopy system performs measurements on one particular absorption band of water; the absorption band around 1450 nm. Due to the influences of stray light, a design of a paper insulation sample holder is proposed. This sample holder is customized to the used components in the design and fixes the samples in place, which limits measurement inconsistencies. A testing procedure is proposed to link the sensor output to the moisture content of the paper insulation sample. A linear relation is found between the sensor output and the moisture content of the paper insulation samples at a moisture content of 0% to 20%.

Traditional search engines rely on centralized databases and powerful servers to process and retrieve information. Developing alternatives to key-value search engine databases in distributed computing environments is a significant challenge, particularly when dealing with limited computational resources. This study explores the use of large language models (LLMs) to address this problem. We focus on environments with constrained computing power, such as mobile devices, to investigate the feasibility of using LLMs as a localized search solution. Through experiments with the state-of-the-art LLMs BERT and T5, we demonstrate their ability to memorize and retrieve unstructured data, specifically YouTube video IDs, based on partial information derived from video titles or tags. Our results show that the explored models can achieve 100\% precision and recall when retrieving 48266 video IDs. The findings suggest that LLMs have the potential to effectively function as a search engine database, offering semantic search capabilities while operating within the constraints of limited computational resources.

There are increasing applications of Visible Light Communication, and LED-to-Camera communication is one of the promising research directions. In this paper, we explore LED-to-Camera communication in mobile scenarios based on the rolling shutter effect to get higher data transmission rate. Movement causes the deformation of the overall shape of the LED in the captured image. And stripes close to the LED edges are too dark to be detected. Both affect the data detection by utilizing the rolling shutter effect. A method is proposed that can increase the transmission rate, which solves the problems caused by movement and darkness of stripes that close to the LED cover edges. After processing, the region of interest is recouped, i.e., the processing range used to detect the rolling shutter stripes becomes larger, which contains more complete data. An approach to detect stripe edges by converting pixel information into binary signal is proposed, which is used to solve the problem of undetectable stripe edges due to high frequency. The proposed method is evaluated through experiments under different conditions. The results demonstrate that the method is effective for addressing the challenges of LED-to-camera communication in mobile scenarios.

With the rapid development of Artificial Intelligence (AI), the size and complexity of models are increasing rapidly. The limited memory and computing power of microcontroller units (MCUs) pose significant challenges for running AI applications on them. This thesis presents a method to run deep learning models on MCUs using a distributed approach.

First, we identified memory size as the primary constraint for implementing deep learning models in MCUs. To address this issue, our method involves splitting the models into smaller weight fragments distributed across multiple networked MCUs. A coordinator MCU manages the overall process, including neuron mapping and data relaying. We demonstrated that our approach reduces peak RAM usage during inference compared to existing methods.

For optimization, we employed layer fusion and quantization to reduce model size while preserving accuracy. We also introduced a rating system to assign capability scores to MCUs for efficient task allocation, explained through mathematical equations.

Our simulation phase validated the effectiveness of our method, showing successful distributed inference in MCUs and providing insights for real-world applications. Implementation on a network of MCUs confirmed the practical applicability and efficiency of our approach.

In conclusion, this thesis presents a feasible and efficient distributed inference method for networked MCUs, addressing resource limitations and enabling practical AI applications on constrained MCU platforms.

In order to have a scalable battery system that can store a lot of charge and output a large amount of power, it is possible to connect multiple battery packs in parallel. This does not come without its own caveats though. During a charge and discharge cycle of a battery pack, the voltage that is present at the output varies. In a scenario where two or more battery packs are connected but have different levels of charge, it is possible that the voltage difference between them is so high that the control system will not allow them to discharge together due to the possibility of large currents flowing between the batteries, which can lead to short and/ or long term damage to the battery. This presents a problem since by only using a percentage of the packs you have connected in parallel, you don’t get the maximum power output or capacity that you could have. The solution to this is to introduce some balancing system. This system will act- ively move charge from one pack to another and allow the voltages to equalise meaning the packs can discharge into a load together again. In this thesis paper, a method for performing this balancing is described, modelled, implemented and verified. It was found that for this specific imple- mentation the time it takes to balance a full battery pack with an empty battery pack is 13 hours and 39 minutes. This full-balance was done by using a PWM signal to limit the current between two packs. Another method was also im- plemented making use of parasitic characteristics of a MOSFET. However, this only served as a method to balance when the battery packs voltages are already somewhat close. Lastly, this thesis provides several fronts for improving the balancing systems and how the implementation can be further fleshed out to shorten the time it takes to balance two packs and how to make it more robust.

Spectrum sensing is a vital technology for alleviating pressure on the radio spectrum and will become more sophisticated as billions more devices come online. In the future, more advanced techniques utilizing deep learning will sense which parts of the spectrum are available to communicate on. This is a heavily researched area, but few papers demonstrate methods of deploying deep learning on the resource-constrained edge devices that will ultimately use them. One approach called Spectrum Painting augments spectrograms and detects which signals are present with a Convolutional Neural Network. We optimize this method for microcontrollers by simplifying the computation of spectrograms and using TinyML techniques. This results in over 90% accuracy on signals with a high signal-to-noise ratio and a latency of 159 ms on a 64 MHz CPU. Our findings conclude microcontrollers are capable of utilizing deep learning for spectrum sensing, but custom hardware will still be required in a real-world deployment.

Visible light positioning (VLP) systems enable indoor positioning through a deployment of light-emitting diodes (LEDs) as transmitters and photodiodes (PDs) as receivers. A promising approach in VLP involves recording the received signal strength (RSS) to construct fingerprint samples for later use in positioning. However, achieving high accuracy demands a labor-intensive data collection process.

In this study, we propose improvements to a data cleaning and augmentation pipeline. Our improvements focus on preserving more source data during cleaning and data-based LED position estimation for more reliable data augmentation. Experimental results show that our approach maintains comparable positioning accuracy while reducing data collection efforts by over 99%. Furthermore, we conduct experiments to investigate the impact of spatially irregular data collection strategies on positioning accuracy. Finally, we deploy a machine learning model on a microcontroller to demonstrate the practical feasibility of our proposed methods.

This paper, in answering the question ”Can effi- cient on-device spectrum sensing be achieved on microcontrollers?”, presents a simple yet compre- hensive approach to signal classification using Con- volutional Neural Networks (CNNs) optimized for deployment on resource-constrained devices. Us- ing data generated via MATLAB’s Wireless Tool- box, as well real world data obtained from testbeds, we created a robust dataset of 9000 samples for training our model. The steps we took while de- veloping a CNN model that performs efficiently on microcontrollers include: data augmentation (pre- processing), model compression and quantization. The model significantly outperformed baseline ac- curacy metrics and maintained competitive infer- ence times, despite the hardware limitations of mi- crocontrollers. This reinforces the idea that Deep Learning has great potential in signal classification. Our research has the potential of being applied to smart homes, IoT networks, industrial automation, and public safety, where our optimized model facil- itates efficient spectrum utilization and minimizes interference.