Human Activity Recognition (HAR) field studies the application of artificial intelligence methods for the identification of activities performed by people. Many applications of HAR in healthcare and sports require the safety-critical performance of the predictive models. The predictions produced by these models should be not only correct but also trustworthy. However, in recent years it has been shown that modern neural networks tend to produce sometimes wrong and overconfident predictions when processing unusual inputs. This issue puts at risk the prediction credibility and calls for solutions that might help estimate the uncertainty of the model’s predictions. In the following work, we started the investigation of the applicability of Non-Bayesian Uncertainty Estimation methods to the Deep Learning classification models in the HAR. We trained a Convolutional Neural Network (CNN) model with public datasets, such as UCI HAR and WISDM, which collect sensor-based time-series data about activities of daily life. Through a series of four experiments, we evaluated the performance of two Non-Bayesian uncertainty estimation methods, ODIN and Deep Ensemble, on out-of-distribution detection. We found out that the ODIN method is able to separate out-of-distribution samples from the in-distribution data. However, we also obtained unexpected behavior, when the out-of-distribution data contained exclusively dynamic activities. The Deep Ensemble method did not provide satisfactory results for our research question. / Inom området Human Activity Recognition (HAR) studeras tillämpningen av metoder för artificiell intelligens för identifiering av aktiviteter som utförs av människor. Många av tillämpningarna av HAR inom hälso och sjukvård och idrott kräver att de prediktiva modellerna har en säkerhetskritisk prestanda. De förutsägelser som dessa modeller ger upphov till ska inte bara vara korrekta utan också trovärdiga. Under de senaste åren har det dock visat sig att moderna neurala nätverk tenderar att ibland ge felaktiga och överdrivet säkra förutsägelser när de behandlar ovanliga indata. Detta problem äventyrar förutsägelsernas trovärdighet och kräver lösningar som kan hjälpa till att uppskatta osäkerheten i modellens förutsägelser. I följande arbete inledde vi undersökningen av tillämpligheten av icke-Bayesianska metoder för uppskattning av osäkerheten på Deep Learning-klassificeringsmodellerna i HAR. Vi tränade en CNN-modell med offentliga dataset, såsom UCI HAR och WISDM, som samlar in sensorbaserade tidsseriedata om aktiviteter i det dagliga livet. Genom en serie av fyra experiment utvärderade vi prestandan hos två icke-Bayesianska metoder för osäkerhetsuppskattning, ODIN och Deep Ensemble, för upptäckt av out-of-distribution. Vi upptäckte att ODIN-metoden kan skilja utdelade prover från data som är i distribution. Vi fick dock också ett oväntat beteende när uppgifterna om out-of-fdistribution uteslutande innehöll dynamiska aktiviteter. Deep Ensemble-metoden gav inga tillfredsställande resultat för vår forskningsfråga.

Modern deep neural networks do not produce well calibrated estimates of their own uncertainty, unless specific uncertainty estimation techniques are applied. Common uncertainty estimation techniques such as Deep Ensembles and Monte Carlo Dropout necessitate multiple forward pass evaluations for each input sample, making them too slow for real-time use. For real-time use, techniques which require only a single-forward pass are desired. Evidential Deep Learning (EDL), and Multiple-Input Multiple-Output (MIMO) networks are prior art in the space of real-time uncertainty estimation. This work introduces EDL-MIMO, a novel real-time uncertainty estimation method which combines the two. The core of this thesis is dedicated to comparing the quality of this new method to the pre-existing baselines of EDL and MIMO alone. / De neurala nätverk vi har idag har svårigheter med att bedöma sin egen osäkerhet utan särskilda metoder. Metoder som Deep Ensembles och Monte Carlo Dropout kräver flera beräkningar för varje indata, vilket gör dem för långsamma i realtid. För realtidstillämpning behövs metoder som endast kräver en beräkning. Det finns redan vetenskapliga artiklar om osäkerhetsmetoder som Evidential Deep Learning (EDL), och Multiple-Input Multiple-Output (MIMO) networks. Denna uppsats introducerar en ny metod som kombinerar båda. Fokus ligger på att jämföra kvaliteten på denna nya metod med EDL och MIMO när de används ensamma / Djúptauganet nútímans eiga erfitt með að meta sína eigin óvissu, án þess að sérstakar óvissumatsaðferðir séu notaðar. Algengar óvissumatsaðferðir líkt og Deep Ensembles, og Monte Carlo Dropout, krefjast þess að djúptauganetið sé reiknað oftar en einu sinni fyrir hvert inntak, sem gerir þessar aðferðir of hægar fyrir rauntímanotkun. Fyrir rauntímanotkun er leitast eftir aðferðum sem krefjast bara einn reikning. Evidential Deep Learning (EDL), og Multiple-Input Multiple-Output (MIMO) networks eru óvissumatsaðferðir sem hafa verið birtar í fyrri greinum. Þessi ritgerð kynnir í fyrsta sinn EDL-MIMO, nýja óvissumatsaðferð sem blandar þeim báðum saman. Kjarni þessarar ritgerðar snýst um að bera saman gæði þessarar nýju aðferðar í samanburð við að nota EDL eða MIMO einar og sér.

This thesis investigates the use of sampled CAD models for training and calibrating a semantic segmentation model, RandLA-Net, with the ultimate goal of localizing modules for digital twinning (the process of creating digital twins). A significant contribution is the development of the Random Placement of Component Generator (RPCG), a synthetic dataset generator that randomly places CAD models within scenes while preserving contextual information such as typical height above ground. Training and testing on datasets generated by RPCG demonstrated its ability to recognize class modules in various randomly generated scenes. Various hyperparameters related to the loss function and pre-processing steps were explored to improve RandLA-Net’s generalization to different contextual settings. Notably, using a class-weighted α in the focal loss showed promise in correctly classifying infrequent classes and reducing network overconfidence under domain shifts with similar prior probability distributions. The semantic segmentation results were promising for the RPCG test set, achieving a mean True Positive Rate (mTPR) of 98% and a mean Intersection over Union(mIoU) of 93.6%. However, the performance on a sampled version of a CAD model representing an installation named Undercentral was comparatively lower, with a mTPR of 41.1% and a mIoU of 33.4%, indicating the need for further adaptation to varied contextual environments. Proposed improvements include enhancing RPCG with an occupancy grid to better simulate compact scenes and evaluating different subsampling rates in RandLA-Net’s random sampling layers. For confidence calibration, the thesis finds that averaging multiple Monte Carlo (MC) dropout evaluations effectively reduces network overconfidence and improves model reliability. Although this work addresses only a portion of the overall digital twinning process, it highlights the potential of synthetic data generation in enhancing semantic segmentation models and contributes towards the localization of modules in digital twin creation.

Knowledge about spatial and temporal variability of hydrological processes is central for sustainable water-resources management, and such knowledge is created from observational data. Hydrologic models are necessary for prediction for time periods and areas lacking data, but are affected by observational uncertainties. Methods for estimating and accounting for such uncertainties in water-resources modelling are of high importance, especially in regions such as Central America. Observational uncertainties were addressed in three ways in this thesis; quality control, quantitative estimation and development of model-evaluation techniques that addressed unquantifiable uncertainties. A first step in any modelling study should be the quality control and concurrent analysis of the representativeness of the observational data. In the characterisation of the precipitation regime in the Choluteca River basin in Honduras, four different quality problems were identified and 22% of the daily data had to be rejected. The monitoring network was found to be insufficient for a comprehensive characterisation of the high spatiotemporal variability of the precipitation regime. Quantitative estimations of data uncertainties can be made when sufficient information is available. Discharge-data uncertainties were estimated with a fuzzy regression for time-variable rating curves and from official rating curves for 35 stations in Honduras. The uncertainties were largest for low flows, as a result of measurement uncertainties and natural variability. A method for calibration with flow-duration curves was developed which enabled calibration to the whole flow range, accounting for discharge uncertainty and calibration with non-overlapping time periods for model input and evaluation data. The method compared favourably to traditional calibration in a test using two models applied in basins with different runoff-generation processes. A post-hoc analysis made it possible to identify potential model-structure errors and periods of disinformative data. Flow-duration curves were regionalised and used for calibration of a Central-American water-balance model. The initial model uncertainty for the ungauged basins was reduced by 70%. Non-representative precipitation data were found to be the main obstacle to comprehensive regional water-resources modelling in Central America. These methods bridged several problems related to observational uncertainties in water-balance modelling. Estimates of prediction uncertainty are an important basis for all types of decisions related to water-resources management. / Kännedom om hur hydrologiska processer varierar i tid och rum är grundläggande för hållbar vattenresursförvaltning och skapas utifrån observerade data. Hydrologiska modeller är nödvändiga för att förutsäga vattenbalansen för tidsperioder och områden utan data, men påverkas av observationsosäkerheter. Metoder för att hantera sådana osäkerheter i vattenresursmodellering är av stor betydelse i regioner såsom Centralamerika. Observationsosäkerheter hanterades på tre olika sätt i denna avhandling; kvalitetskontroll, kvantitativ uppskattning och utveckling av modellutvärderingsmetoder för beaktande av icke kvantifierbara osäkerheter. Ett viktigt första steg är kvalitetskontroll och samtidig analys av datas representativitet. Vid karaktäriseringen av nederbördsregimen i Cholutecaflodens avrinningsområde i Honduras identifierades fyra olika kvalitetsproblem och 22 % av data sorterades bort. Stationsnätet var otillräckligt för en fullödig karaktärisering av nederbördsregimens variationer i tid och rum. Dessa var mycket stora som ett resultat av komplexiteten hos de nederbördsgenererande mekanismerna. Kvantitativ uppskattning av observerade datas osäkerhet kan göras när tillräcklig information är tillgänglig. Osäkerheter i vattenföringsdata uppskattades dels vid beräkning av vattenföring med en oskarp regression för en tidsvariabel avbördningskurva, dels från en analys av officiella avbördningskurvor från 35 stationer i Honduras. Osäkerheten var i båda fallen högst vid låga flöden som ett resultat av högre mätosäkerheter samt större naturlig variabilitet än vid höga flöden. En metod för modellkalibrering med varaktighetskurvor utvecklades och gjorde det möjligt att kalibrera för hela flödesintervallet samtidigt, ta hänsyn till osäkerheter i vattenföringsdata samt kalibrera med icke överlappande driv- och utvärderingsdata. Metoden testades med två olika modeller i två avrinningsområden med olika avrinningsbildningsprocesser, och visade goda resultat jämfört med traditionell modellkalibrering. En post hoc-analys gjorde det möjligt att identifiera troliga modellstrukturfel och perioder med disinformativa data. Varaktighetskurvor regionaliserades och användes för kalibrering av en regional vattenbalansmodell för Centralamerika, varvid den initiala modellosäkerheten minskades med 70 %. Icke representativa nederbördsdata identifierades som det största hindret för regional vattenresursmodellering i Centralamerika. De metoder som utvecklades i detta arbete gör det möjligt att överbrygga ett flertal problem orsakade av bristfällig tillgänglighet och kvalitet av data och leder därmed till en förbättrad uppskattning av osäkerheten i vattenbalanssimuleringar. Sådana osäkerhetsskattningar är ett viktigt underlag vid alla typer av förvaltningsbeslut som rör vattenresurser.

The automotive safety company Veoneer are producers of high end driver visual assistance systems, but the knowledge about the absolute accuracy of their dynamic calibration algorithms that estimate the vehicle’s orientation is limited. In this thesis, a novel measurement system is proposed to be used in gathering reference data of a vehicle’s orientation as it is in motion, more specifically the pitch and roll angle of the vehicle. Focus has been to estimate how the uncertainty of the measurement system is affected by errors introduced during its construction, and to evaluate its potential in being a viable tool in gathering reference data for algorithm performance evaluation. The system consisted of three laser distance sensors mounted on the body of the vehicle, and a range of data acquisition sequences with different perturbations were performed by driving along a stretch of road in Linköping with weights loaded in the vehicle. The reference data were compared to camera system data where the bias of the calculated angles were estimated, along with the dynamic behaviour of the camera system algorithms. The experimental results showed that the accuracy of the system exceeded 0.1 degrees for both pitch and roll, but no conclusions about the bias of the algorithms could be drawn as there were systematic errors present in the measurements. / Bilsäkerhetsföretaget Veoneer är utvecklare av avancerade kamerasystem inom förarassistans, men kunskapen om den absoluta noggrannheten i deras dynamiska kalibreringsalgoritmer som skattar fordonets orientering är begränsad. I denna avhandling utvecklas och testas ett nytt mätsystem för att samla in referensdata av ett fordons orientering när det är i rörelse, mer specifikt dess pitchvinkel och rollvinkel. Fokus har legat på att skatta hur osäkerheten i mätsystemet påverkas av fel som introducerats vid dess konstruktion, samt att utreda dess potential när det kommer till att vara ett gångbart alternativ för att samla in referensdata för evaluering av prestandan hos algoritmerna. Systemet bestod av tre laseravståndssensorer monterade på fordonets kaross. En rad mätförsök utfördes med olika störningar introducerade genom att köra längs en vägsträcka i Linköping med vikter lastade i fordonet. Det insamlade referensdatat jämfördes med data från kamerasystemet där bias hos de framräknade vinklarna skattades, samt att de dynamiska egenskaperna kamerasystemets algoritmer utvärderades. Resultaten från mätförsöken visade på att noggrannheten i mätsystemet översteg 0.1 grader för både pitchvinklarna och rollvinklarna, men några slutsatser kring eventuell bias hos algoritmerna kunde ej dras då systematiska fel uppstått i mätresultaten.

Artificial intelligence (AI) has been seeing a great amount of hype around it for a few years but more so now in the field of diagnostic medical imaging. AI-based diagnoses have shown improvements in detecting the smallest abnormalities present in tumors and lesions. This can tremendously help public healthcare. There is a large amount of data present in the field of biomedical imaging with the hospitals but only a small amount is available for the use of research due to data and privacy protection. The task of manually segmenting tumors in this magnetic resonance imaging (MRI) can be quite expensive and time taking. This segmentation and classification would need high precision which is usually performed by medical experts that follow clinical medical standards. Due to this small amount of data when used with machine learning models, the trained models tend to overfit. With advancing deep learning techniques it is possible to generate images using Generative Adversarial Networks (GANs). GANs has garnered a heap of attention towards itself for its power to produce realistic-looking images, videos, and audios. This thesis aims to use the synthetic images generated by progressive growing GANs (PGGAN) along with real images to perform segmentation on brain tumor MRI. The idea is to investigate whether the addition of this synthetic data improves the segmentation significantly or not. To analyze the quality of the images produced by the PGGAN, Multi-scale Similarity Index Measure (MS-SSIM) and Sliced Wasserstein Distance (SWD) are recorded. To exam-ine the segmentation performance, Dice Similarity Coefficient (DSC) and accuracy scores are observed. To inspect if the improved performance by synthetic images is significant or not, a parametric paired t-test and non-parametric permutation test are used. It could be seen that the addition of synthetic images with real images is significant for most cases in comparison to using only real images. However, this addition of synthetic images makes the model uncertain. The models’ robustness is tested using training-free uncertainty estimation of neural networks.

This thesis investigates predicting the Intersection over Union (IoU) in detections made by the face detector YOLO5Face, which is done to use the predicted IoU as a new uncertainty measure. The detections are done on the face dataset WIDER FACE, and the prediction of IoU is made by adding a parallel head to the existing YOLO5Face architecture. Experiments show that the methodology for predicting the IoU used in this thesis does not work and the parallel prediction head fails to predict the IoU and instead resorts to predicting common IoU values. The localisation confidence and classification confidences of YOLO5Face are then investigated to find out which confidence measure is least uncertain and most suitable to use when identifying faces. Experiments show that the localisation confidence is consistently more calibrated than the classification confidence. The classification confidence is then calibrated with respect to the localisation confidence which reduces the Expected Calibration Error (ECE) for classification confidence from 0.17 to 0.01.