Accurate pedestrian navigation in Global Navigation Satellite System (GNSS)-denied environments remains challenging. Pedestrian positioning systems should be self-contained and constrained by size, weight, and power. In some applications, such as defence, these systems should additionally be passive and difficult to detect. Inertial Measurement Unit (IMU)-only navigation is attractive in this setting, but direct integration of inertial measurements leads to rapid drift due to sensor noise and bias. Recent data-driven IMU-only pedestrian inertial-odometry approaches reduce this drift by using neural networks to provide short-term motion pseudo-measurements that are fused with inertial propagation in a filtering framework. However, human locomotion is heterogeneous, and a single model trained on all locomotion types may not perform equally well across different motion types.

This thesis investigates how estimated locomotion-type information can be incorporated into data-driven IMU-only pedestrian inertial odometry to improve state-estimation accuracy and uncertainty consistency. The proposed motion-aware learned-filter method combines a Convolutional Neural Network (CNN)-based locomotion-recognition module, motion-specific velocity-regression networks, and an Extended Kalman Filter (EKF)-based fusion architecture. The locomotion-recognition module classifies the current locomotion type from short IMU windows. The network corresponding to that class then predicts a body-frame velocity pseudo-measurement and its associated uncertainty for the EKF update. Two types of method variants are also investigated. The first uses the full probability distribution over locomotion classes in the EKF update. The second adapts the training settings separately for each locomotion class.

The results show that locomotion type can be recognized reliably from short IMU windows, although the classifier confidence decreases during transition periods. The motion-specific velocity-regression models improve neural-network prediction accuracy for selected locomotion types, especially running, sideway walking, and stairs up and down. The clearest improvement after EKF-based fusion is observed for running, where the average trajectory root mean square error (RMSE) over the tested trajectories decreases from 10.328 m to 8.063 m, corresponding to an improvement of approximately 22%. In some cases, such as sideway walking, selecting the motion-specific model using the ground-truth locomotion label does not necessarily produce the best trajectory estimate.

The investigated motion-specific training adaptations further reduce prediction errors for some motion-specific velocity-regression models, although the magnitude of the improvement depends on the locomotion type. The method variants that use the full probability distribution over locomotion classes have only a marginal effect when a single isolated transition occurs. However, on the dataset with more frequent locomotion changes, combining predictions from all motion-specific models reduces the average trajectory RMSE by approximately 6% compared with the proposed motion-aware learned-filter method. ... Read more

Results show that combined dual-sensor features consistently outperform single-sensor variants, and that coarser grids yield higher exact accuracy while mean physical error remains stable across resolutions at approximately 70–80 cm. At 0.5 m resolution, the best configuration places 71% of predictions within 50 cm of the true location, approaching the accuracy of a UWB baseline system while requiring no installed infrastructure. Zone merging and hexagonal tessellation do not provide consistent improvements over the plain square grid, suggesting that magnetic ambiguity rather than data imbalance or cell geometry is the dominant source of error. The findings demonstrate that infrastructure-free magnetic fingerprinting is a practically viable approach for coarse spatial awareness in socially dynamic indoor environments. ... Read more

Controllers currently used by aerospace vehicles such as helicopters and aircraft are able to reject almost completely the disturbances introduced by wind in their attitude and position. Once attenuated, wind gusts just create some random, low amplitude oscillations in the vehicle, whose more severe consequence in common flight scenarios is a reduction in crew and passengers comfort. However, for certain high-precision operations, these oscillations can put the mission at risk. This is the case for rescue helicopters, which need to hover over precise spots in the sea or mountains to allow personnel to assist injured people in difficult-access areas. In many occasions, the accuracy with which helicopters are able to hold a desired position in adverse weather conditions is not enough to safely perform a rescue mission.
The disturbance rejection performance of helicopters can be improved by providing their controllers with angular acceleration information of the vehicle. Although angular acceleration sensors are available on the market, their current cost and weight prevent their widespread use.
This master thesis covers the design and testing of a novel estimation method for angular acceleration in aerospace vehicles based on linear acceleration measurements. The proposed estimator is a Stationary Kalman Filter with Modified Input (MISKF), which outperforms current approaches found in industry in terms of accuracy and time delay, while being lightweight and Linear-Time-Invariant (LTI). This observer relies only on gyroscopes and linear accelerometers, which are usually present in most vehicles. In addition to this, it does not require the dynamic model of the system in which it is used, nor the accurate position of its Center of Gravity (C.G.). These advantages make the MISKF applicable in a broad range of aerospace vehicles, including helicopters. The proposed observer is compared to other angular estimation methods such as speed differentiators using real flight data, showing a faster convergence and lower noise amplification, as well as to other observers like an Extended Kalman Filter (EKF), leading to similar performance while being more compact and lightweight. ... Read more

Indoor localization is used in applications such as emergency response, healthcare, and navigation in large indoor spaces. Global Navigation Satellite System (GNSS) is widely used outdoors but is unreliable indoors due to signal obstructions. In addition, many indoor localization methods depend on dedicated infrastructure. As an infrastructure-free alternative, handheld smartphone Pedestrian Dead Reckoning (PDR) uses inertial sensors but accumulates drift over time. This thesis reduces that drift by integrating handheld smartphone PDR with one-dimensional magnetic-field Simultaneous Localization and Mapping (SLAM), which jointly estimates the pedestrian trajectory while building a map from sensor measurements. It uses a one-dimensional magnetic map along the walked path, which is more lightweight than constructing full two-dimensional or three-dimensional magnetic maps. The drift is reduced by detecting magnetic-field loop closures, i.e., revisits to previously traversed locations. First, a handheld smartphone PDR trajectory is estimated from detected steps, estimated step lengths, and headings obtained from inertial sensing and orientation filtering. The PDR provides the dynamical model for the Kalman Filter (KF) position estimate update. To correct drift, three-axis magnetometer measurements are used to detect loop closures by matching magnetic sequences along the walked path. Magnetic similarity is combined with the KF positional uncertainty to score candidate loop closures. Candidates are validated using temporal constraints, a magnetic-field excitation condition, and a marginal-likelihood test based on the KF innovation. Accepted loop closures are incorporated as landmark constraints by augmenting the KF state, yielding a corrected trajectory. The method is evaluated on three real-world handheld smartphone datasets collected inside a university building over approximately 500 m. In addition to standard walking, the evaluation includes PDR failure modes such as heading drift and step-length underestimation to test robustness, and shows that one-dimensional magnetic-field SLAM can reduce their impact. Four magnetic sequence-matching loop-closure detection methods are compared; the results indicate that using magnetic-field sequences downsampled to footsteps provides the best loop-closure detection. For the nominal walk, magnetic-field SLAM reduces the Absolute Trajectory Error from 5.5 m to 1.8 m and the final error from 11.0 m to 1.0 m. ... Read more

Magnetic-aided navigation using unmanned aerial vehicles (UAVs) is a promising method in case traditional navigation methods fail, but aeromagnetic platform noise from electric motors, electronics, and actuators can mask subtle geological signals. Traditional compensation methods, such as the Tolles-Lawson (TL) model, assume linear relationships between platform orientation and platform noise, failing to capture complex, time-varying disturbances from dynamic onboard systems. Existing machine learning approaches typically require noise-free reference measurements or known anomaly maps, resources often unavailable in practical surveying scenarios.

This thesis develops data-driven, reference-free methods for compensating platform noise in aeromagnetic measurements. The research addresses two key questions: whether deep learning methods can effectively predict and compensate platform noise without ground-truth references, and which drone subsystems contribute most significantly to platform noise.

A validation apporoach was implemented using flight data where crustal anomalies are naturally attenuated, enabling reference-free performance assessment. Comprehensive data from a fixed-wing UAV equipped with scalar and vector magnetometers, logging 273 platform-related input signals during figure-of-merit manoeuvres over an 800~m $\times$ 800~m survey area was used.

The compensation approach employed hierarchical modelling: Extended Tolles-Lawson (ETL) compensation incorporating drone inputs projected onto the magnetic field direction, followed by multilayer perceptron (MLP) neural networks trained to predict residuals. SHAP (SHapley Additive exPlanations) analysis provided model-agnostic feature importance assessment to identify the most influential platform inputs.

Results demonstrate substantial improvements over traditional methods. ETL compensation achieved improvement ratios averaging 7.31 for scalar magnetometers and 18.76 for vector magnetometers, compared to 4.90 and 10.19 respectively for standard TL compensation. The combined ETLNN approach (ETL + neural network) further enhanced performance to average improvement ratios of 8.87 on total field magnetometer and 22.22 on vector magnetometer data, a significant improvement over traditional TL methods.

SHAP analysis revealed that engine-related parameters (battery current, throttle commands), inertial measurement data (accelerations, gyroscopic rates, vibration), and attitude information (roll, pitch, yaw) are the primary contributors to platform noise. Features projected onto the magnetic field direction consistently outperformed raw inputs, validating the physical basis for this transformation, while derivative features contributed minimally whilst increasing overfitting.

The primary limitation is the inability to validate performance on data containing actual magnetic anomalies, as the high-altitude validation approach deliberately suppressed geological signals. Future work should prioritise validation using artificial magnetic sources or reference magnetometer configurations to assess preservation of genuine geological signals.
... Read more

Many applications such as indoor navigation, search and rescue operations, and surveying, require accurate localization in challenging environments. In environments where absolute positioning (using for example GNSS receivers) fails due to dense obstruction of the required signals, positioning algorithms rely solely on dead reckoning, often using measurements from an Inertial Measurement Unit (IMU). This dead reckoning process suffers from integration drift due to the accumulation of errors in the sensor measurements. Pedestrian Dead-Reckoning (PDR) algorithms can reduce this integration drift in cases where the IMU is held by a walking person, by leveraging patterns from the periodic walking motion. This thesis investigates a state-of-the-art PDR algorithm from Liu et al. that combines an Extended Kalman Filter (EKF) with a velocity-predicting neural network that corrects the filter in the measurement update to mitigate integration drift. The focus is on finding if its performance can be improved by adapting the neural network in the algorithm. First, adaptations of the network’s parameters have been experimented with to investigate their effect on the algorithm’s accuracy and search for a bottleneck that limits it. This bottleneck was found to be a bias in the predicted velocity by the network, as this violates the EKF assumption that the error in the measurement update is distributed with zero mean. The second part investigates how using ideas from the Physics-Informed Neural Network (PINN) as an alternative to the data-driven neural network in the PDR algorithm affects its performance. Four network architectures have been trained using a loss function that includes a penalty for errors in the physics of the predicted velocities of the system. Training these PINN-inspired networks required a much longer training time. The results show that using this physics loss does not show significant improvements in the accuracy of the PDR algorithm, as the bias in the velocity prediction is not addressed by the physics loss. This thesis concludes that the main limitation of the neural network in the PDR algorithm by Liu et al. is a bias in the predicted velocity, and that a PINN is unable to provide significant improvements in the accuracy of the algorithm, whilst costing much more computational resources to train. It is therefore not recommended to use a PINN in this PDR algorithm, and the optimal configuration of the network was found to be a slightly adapted version of the network used by Liu et al., using IMU and device tilt data sampled at 50 Hz as network input. Future research could focus on understanding the origin of the bias in velocity predictions and to mitigate it once its source is known. ... Read more

Ferromagnetic materials in the walls and ground of buildings cause perturbations in the earth’s magnetic field. These spatial variations in the ambient magnetic field observed in indoor environments are mostly time-invariant. The deviations of the magnetic field can be captured in a magnetic field map, providing valuable information for indoor localization and navigation. Gaussian processes are a useful tool to model the ambient magnetic field, where the characteristics of the magnetic field are specified by the hyperparameters of the Gaussian process. To avoid the computational difficulties associated with full Gaussian process regression, a reduced-rank approximation is implemented. The magnetic field is measured using a magnetometer. A motion capture suit contains multiple magnetometers, whose relative positions are accurately determined. In this thesis it is researched if the quality of the magnetic field map can be improved when the measurement come from the motion capture suit. This is a challenging task, since the magnetometers are placed at different altitudes and the characteristics of the indoor magnetic field also vary with altitude. Consequently, the hyperparameters that best fit the measurement data are different per magnetometer. The negative effects of this observation are reduced by considering subsets of magnetometers operating at similar altitudes. However, an evident relation between the magnetometer’s altitude and its optimal hyperparameters has not been found. The hypothesis that complementing measurements from a single magnetometer with measurements of magnetometers operating at similar altitudes would improve the quality of the magnetic field is not supported based on the experiments. It is expected that either the magnetometer measurements lacked consistency, or that the assumption that the MOCAP suit returned the true magnetometer locations does not apply.
... Read more

The localization and tracking of artists on stage enables theatre spotlights to automatically follow the artist’s movements. The company Sendrato utilizes Ultra-Wideband (UWB) systems for this purpose, but the position estimation accuracy decreases when the sensors operate in a Non-Line-of-Sight (NLOS) environment. To prevent the body from blocking the UWB signals, Sendrato places two UWB tags on the hips of the artist and averages the two estimated tag positions. Inertial Measurement Units (IMUs) offer an additional means for location tracking, providing inputs to calculate position, velocity, and orientation estimates independent of the environment, but suffer from accumulative error due to integration drift. This thesis studies how Sendrato can improve position tracking accuracy with UWB data by incorporating IMU sensor data. Additionally, the thesis investigates how the two tags can be coupled to correct for each other’s inaccuracies. Hence the thesis studies how the sensor fusion of two coupled UWB/IMU sensors, attached to a person’s hips, can be used to improve position tracking accuracy compared to using two separate UWB tags. This gives rise to a research approach consisting of two parts. To investigate the potential of UWB/IMU sensor fusion for position tracking accuracy, an Extended Kalman Filter (EKF) fusing IMU and UWB measurements is implemented and compared to UWB-only position tracking algorithms. Additionally, it is investigated whether IMU bias state estimation, Zero Velocity Update (ZUPT) implementation and NLOS detection and mitigation can further improve the UWB/IMU EKF tracking accuracy. The second part researches the potential of coupling two UWB/IMU tags for position tracking accuracy. Previous methods inspired to use knowledge of the fixed relative distance between the tags to correct position estimates from each tag. This research developed this approach by including an equality constraint on the distance between the tags into the EKF. An experiment is conducted where the joint UWB/IMU EKF is tested on a known walked trajectory containing several stationary points. The results show that the tightly coupled UWB/IMU EKF can help smooth faulty UWB measurements, correct stationary points with a ZUPT, and identify IMU bias. Moreover, enforcing a fixed distance between joint tags allows for mutual correction of their trajectories, though the impact on the averaged trajectory may be less significant. All of these techniques show potential for improving position tracking accuracy compared to a loosely coupled UWB-only algorithm. However, the methods all showed limitations, presumably caused by the data quality. An important direction for future work would be to continue this research with better calibrated UWB data.
... Read more

IMU-to-Segment(I2S) calibration is a critical step in using IMU for human motion capture, as it determines the relative orientation and position between the IMU and the body segment it is attached to. Traditional constraint-based method rely on kinematic constraints to perform I2S calibration but fail in scenarios where no relative motion occurs between connected segments. To address this limitation, this thesis extends existing deep learning approaches for I2S calibration to the stiff case and investigates methods to further enhance calibration accuracy.

After completing I2S calibration using deep learning with a single IMU in stiff case, this thesis further explores joint training with dual-IMU model and integrates kinematic constraints into the model. Experiment results demonstrate that the joint training allows the models to leverage inter-IMU motion information, improving the model performance. Furthermore, integrating kinematic constraints with appropriate weights into the loss function of deep learning model improves calibration accuracy by guiding predictions to satisfy physical constraints. However, overly large constraint weights may result in larger calibration error.

This thesis provides insights into how deep learning can be adapted to address the challenges of I2S calibration in stiff joint scenarios. It also combines deep learning with kinematic information through joint training and the integration of kinematic constraints, achieving improved calibration accuracy. Future work will explore the application of this approach to real-world motion data and the integration of diverse kinematic constraints. ... Read more

Indoor localization is an active research area since traditional localization methods, such as global navigation satellite systems (GNSS), are ineffective in indoor environments. One promising solution is to make use of anomalies in the magnetic field caused by ferromagnetic materials within buildings. By constructing a magnetic field map in an indoor location, the spatial variability in the magnetic field can be utilized to aid in inferring the location. Gaussian process regression (GPR) is frequently used to model the magnetic field and provide magnetic field predictions as well as the corresponding predictive uncertainty. In addition to mapping an entire building or room, the existing literature has also been studying mapping in close proximity to the sensor, as this provides valuable information for odometry purposes. Magnetic field measurements are obtained using sensors known as magnetometers. Magnetometer arrays, which are typically two- or three-dimensional fixed structures that contain multiple magnetometers, are used to construct these local maps as they are capable of measuring the magnetic field at multiple locations simultaneously. However, the time complexity of GPR scales cubically with the number of data points. When these magnetometer arrays are used, the size of the data increases rapidly. As a result, GPR can quickly become computationally intractable in combination with arrays. In this thesis, building on prior work that demonstrated the use of spatial derivatives of arrays for odometry purposes, we instead analyze how effective the information on the array can be approximated using the spatial derivative of the magnetic field to alleviate some of the computational burden. We analyze this by examining the performance of using the spatial derivative computed from different array configurations in terms of the number of magnetometers, their spacing, and noise levels, and comparing it with using all magnetometer measurements on the array. Through simulations, we show that the spacing is a key factor in determining the spatial derivative, and that more magnetometers and less noise result in better estimates. It is shown that the spatial derivative is not an effective approximation to the full information on the array. At least for the arrays considered in this thesis, it did not reduce the computational burden sufficiently to justify the resulting loss in map quality. In addition to examining which array configurations perform best with spatial derivatives, we also evaluate which configurations yield accurate maps for existing odometry methods that rely on arrays. We show that the array configuration plays an important role and requires deliberate selection for optimal performance.
... Read more

Nonlinear partial differential equations (PDEs) are generally hard to solve, but over the past 65 years the notions of Lax integrability and nonlinear Fourier transforms (NFTs) have been developed to solve a large class of so called Lax-integrable partial differential equations. If we are able to find a suitable pair of linear operators, consisting of a spectral operator and a propagation operator, a so-called Lax pair, we are able to analytically solve the nonlinear PDE.If this pair of linear operators is a correct Lax pair for the PDE, then the spectral operator can be used to define an NFT, which transforms the signal from the physical domain to a PDE-specific spectral domain, in which all spectral components are independent, similar to the regular Fourier transform for linear problems.
In the spectral domain, the transformed signal can be propagated using the linear propagation operator, after which the propagated signal can be transformed back into the physical domain, hence solving the nonlinear PDE.

The presence of a Lax pair can thus be used to solve a nonlinear PDE, but it also yields insight into the underlying dynamics of the system. Therefore, the identification of a Lax pair for a system is of great value. However, there is no general method to find a Lax pair given an evolution equation. Systems may also be non-integrable, and therefore a corresponding Lax pair may not even exist. To the best of our knowledge, no practical general methods are known in the literature to determine whether or not a system is Lax-integrable. Since the general problem is very complex, one may focus on more specific cases for practical purposes, in which a Lax-integrable PDE is sought using measurement data.

In this thesis, we therefore focus on data-driven identification of Lax-integrable partial differential equations (PDEs), specifically those of the AKNS-type. The primary objective is to find a Lax-integrable PDE that best explains given experimental measurement data, enabling a comprehensive analysis using the nonlinear Fourier transform (NFT). While real-world systems may not be exactly Lax-integrable due to imperfections, many are well approximated by such equations, including scenarios such as fiber optical wave propagation, surface wave propagation in shallow water canals, and mechanical wave propagation in coupled pendulums. ... Read more

Probabilistic or Bayesian modeling plays a fundamental role in engineering and science, providing a framework for integrating noisy measurements with predictive models through probability distributions.
While probabilistic methods have many benefits, such as recursive estimation and uncertainty quantification, they often come with substantial memory and compute requirements.
Computational challenges are particularly pronounced in large-scale settings, where data sets contain a high number of measurements, and for high-dimensional problems, which require exponentially many parameters to describe probability distributions.
These scenarios can suffer from the curse of dimensionality, which requires exponentially growing computing resources, making conventional approaches computationally intractable.

This dissertation addresses computational challenges by leveraging tensor networks (TNs) to develop computationally efficient probabilistic algorithms.
TNs, also known as tensor decompositions, extend matrix decomposition to higher dimensions by representing large multidimensional arrays, i.e., tensors, in a compact, decomposed format, defined by TN components and TN ranks.
Under the assumption of low-rank structure, TNs enable efficient storage and computation, making large-scale and high-dimensional problems more tractable, even on resource-constrained hardware such as conventional laptops.
The focus of this work is on scalable solutions for Bayesian estimation problems involving Gaussian distributions and exact inference, including recursive filtering and Gaussian process (GP) regression. ... Read more

Accurate ship motion prediction is crucial for safe and efficient operations at sea. Previously studied Single Input Single Output (SISO) system-based method often falls short in providing robust and precise predictions. This thesis explores the potential of using a Multiple Inputs Multiple Outputs (MIMO) system identification approach to provide more accurate and ro- bust ship motion prediction in offshore operations.
We investigate two MIMO systems, the force-to-motion system and wave-to-motion system, employing Spectral Analysis method and Subspace method, respectively. Both synthetic and real-field data are utilized to study the methods. While Spectral Analysis method demon- strates better accuracy, it relies on the assumption of perfect accuracy of the pre-computed force Response Amplitude Operators (RAOs), which is a limitation of this approach. Sub- space method is less precise, it also requires real-time re-evaluation of the system order due to its sensitivity to wave conditions, making this method less versatile.
Comparing these methods with SISO system method and conventional RAO-based method, Spectral Analysis method emerges as the most accurate, followed by Subspace method. These findings strongly support the potential of MIMO system identification to significantly improve ship motion prediction accuracy.
... Read more

Magnetometers are widely equipped in smartphones. They measure the direction and the magnitude of the magnetic field of the environment. Since the measurements are not transition data, there is no drift when estimating position and orientation using a magnetometer. Furthermore, magnetic field localization using magnetometers requires no extra devices set in the environment, and this indicates the cost of localization using a magnetometer can be lower than other localization methods that need multiple devices set in the localizing area. Therefore, magnetic field localization is an interesting method for indoor localization. However, there exists a research gap in the algorithms that have been applied to magnetic field localization. In the current research, the Extended Kalman filter (EKF) and the Particle filter (PF) are applied to magnetic field localization. The EKF is more efficient than the PF, but has low accuracy when the distribution is multimodal. On the other hand, the PF is more computationally costly compared to the EKF but is more robust to the multimodality. As a result, a survey of the possible solutions to the current research gap was carried out. From this survey, Gaussian sum filter (GSF) was found to be a promising candidate as the solution to the research gap. To test the performance and assumptions of the GSF, the GSF was applied to a fully simulated magnetic field localization system and a localization system with the measurements obtained from a real-world magnetometer. The results from these simulations show that the GSF is more suitable for multimodality than the EKF. Besides, the computational cost of the GSF is found to be lower than the PF while the GSF has an equivalent or even better accuracy than the PF. ... Read more

This thesis scales methods for Gaussian process-based magnetic field mapping and localization
in five distinct ways. ... Read more

Drifting, a specialized form of sideslip control, involves intentionally inducing and maintaining a state of oversteer for lateral sliding of the vehicle. While previous research has primarily focused on autonomous drift control, the integration of the driver in the control loop remains largely unexplored. This thesis aims to explore how a vehicle sideslip control system can be designed and evaluated for driver-involved drifting scenarios. To this end, a seven degree-of-freedom vehicle model and a tire model calibrated on data from a test vehicle were used to develop a model-based optimization scheme, leveraging the capabilities of all-wheel torque vectoring to achieve the desired drifting behavior. Phase portrait and real data analyses during drifting maneuvers were conducted to understand drift dynamics and develop the reference generator, which calculates targets for the controller. The reference generator and controller were validated through simulations using IPG CarMaker, drifting along a constant radius circle and a section of a track. The results demonstrate that the developed control system enables stable and controllable drift behavior with a driver-in-the-loop and allows drivers to fully exploit their vehicle’s capabilities. ... Read more

Inertial Measurement Units (IMUs) have become increasingly popular human motion estimation due to their portability, self-contained features, and cost-effectiveness compared to marker-based sensing systems which rely on external cameras to observe the position of the markers. Over the years, many studies have proposed various models and algorithms based on IMUs and the kinematics of the human body for motion estimation, neglecting the fact that IMUs measure acceleration, which can be directly related to joint torque with known inertial parameters. In turn, by estimating joint torque and incorporating kinetics into the model, it becomes possible to address a long-standing problem in the field of biomechanics: the marker-based sensing system’s inability to provide a reliable estimation of kinetics due to the need for numerical differentiation. In a previous study, a kinetics model was proposed for estimating the motion but validated only on a robotic arm. In our study, we further explore the performance of using IMUs to estimate wheelchair user motion data based on Extended Kalman Filter (EKF) and the kinetics model, with marker-based Inverse Kinematics (IK)/Inverse Dynamics (ID) as the benchmark.

Compared to Marker-based IK, the method leveraging kinetics achieves a Root Mean Squared Difference (RMSD) below 16◦ for three out of four tasks across all joints throughout the trial. After analyzing the only task with degradation in estimation, we conclude that the erroneous IMUs measurements results from Soft Tissue Artefacts (STA) is the most likely reason. For joint torque estimation, the RMSD for joint torque estimation is below 3.05Nm for the tasks less affected STA. Through fine-tuning the EKF, we can achieve fast and responsive estimation results without being affected by numerical differentiation, enabling us to capture both sudden and subtle changes in joint torque estimation. The kinetics model performs better than the kinematics-based model in estimating both kinematics and kinetics and also reduced drifting behavior. Compared to OpenSense, which depends on magnetometer measurements, kinetics model estimation shows comparable kinematics estimation accuracy while excluding
the use of heading information. The results show that including the kinetics model for human motion estimation can improve estimation accuracy and robustness encouraging further studies to include kinetics for human motion estimation. ... Read more

Loop Robots develops and operates the next generation of fully autonomous disinfection robots in hospitals and healthcare settings. Accurate localization is essential in order to navigate reliably and effectively disinfect the tight hallways and corners of a patient room, operating room, or intensive care unit in a hospital. Odometry, or dead-reckoning, is the process of estimating the robot’s pose with respect to a known initial pose. It forms the backbone of the robot’s localization strategy, as well as being critical to many other processes that run on the robot, such as control and mapping.

The current strategy of generating odometry using the wheel encoders is prone to drift due to the integration of errors into the estimate. Moreover, the estimation takes place on the horizontal plane due to the nature of the wheel encoders. To improve the odometry and extend it to the full 3D space, a solution that makes use of all of the onboard sensors in a tightly coupled manner is required.

In this thesis, we make the first steps towards this larger goal. The contributions of this thesis include: (1) an Extended Kalman Filter (EKF) algorithm to fuse data from the Inertial Measurement Unit (IMU) and wheel encoders to estimate position and orientation of the robot in 6-DoF; (2) a line feature extraction and tracking methodology to extract primitives from LiDAR data; and (3) a Moving Horizon Estimation (MHE) scheme based on a factor-graph formulation to perform pose estimation on the horizontal plane using LiDAR data and wheel encoders.

We test the three modules individually using a combination of simulations and real-world data wherever possible. We found that the MHE scheme is able to reduce drift over the long term, but is sensitive to the effects of outliers in feature matching, motion distortion of LiDAR scans, and wheel slip. The EKF scheme is able to reduce the overall drift and correct for wheel slips.

Based on these results, promising avenues for the improvement of all the proposed modules are given, along with recommendations on how to combine them in a tightly coupled fashion. ... Read more

Artificial Neural Networks (ANNs) have emerged as a powerful tool for classification tasks due to their ability to outperform traditional methods. Nevertheless, their effectiveness relies heavily on the availability of large, varied, and labeled datasets, which are often not available. To counter this constraint, data augmentation techniques have emerged, leveraging existing data to generate additional, variant data. Extending these techniques to multi-dimensional time series data, such as the transportation mode detection data considered in this thesis, however, introduces challenges. In response, generative models such as Variational Autoencoders (VAEs) have shown promising advancements.

In this context, this thesis investigates the application of the Iterative Hierarchical Data Aug- mentation (IHDA) algorithm for ANNs, which represents a VAE-based data augmentation technique. The IHDA method utilizes VAEs not only to generate new data samples but also to map existing data to a lower-dimensional latent space, which is then utilized for identifying samples that might require additional training. The proponents of this method, Khan and Fraz, reported an accuracy elevation for the considered transportation mode detection classifier from 83% to 92%. However, due to the absence of publicly accessible code for this algorithm, the initial step of this thesis involved implementing the IHDA algorithm. Further, this research proposed and incorporated advancements like the σ-VAE, aimed to improve the generative capacity of the VAE and to refine its latent space mapping. Additionally, the Kullback-Leibler (KL) divergence was introduced as a similarity metric, aiming to optimize the identification process of samples that require retraining.

Unfortunately, the results reported by Khan and Fraz could not be reproduced in this study. Furthermore, despite the potential shown by the σ-VAE to improve the generative capacity and refine the latent space mapping, along with the enhanced sample identification through the KL divergence, these enhancements did not lead to an overall improvement in the IHDA algorithm. This was primarily attributed to the low generative performance of the VAEs utilized, which also hindered a thorough evaluation of the effectiveness of the IHDA algorithm.

Given these outcomes, it is suggested that future work should focus on employing more complex VAE models with the potential to enhance their generative performance, which, in turn, could improve the IHDA algorithm’s overall effectiveness. ... Read more

This thesis investigates the performance of the invariant extended Kalman filter (IEKF) compared to the multiplicative extended Kalman filter (MEKF) in the context of nonlinear state estimation on matrix Lie groups. The IEKF, a relatively recent variant of the EKF, is particularly suitable for systems with group-affine process models and invariant measurement models. When applied to such systems, the IEKF exhibits guaranteed state-independent error dynamics, which proves advantageous in cases of poor or inaccurate system initialization.

While previous studies have highlighted the benefits of the IEKF in poorly initialized systems, it is unclear whether the IEKF and the multiplicative EKF exhibit significant differences in performance when the system is already accurately initialized. Therefore, this thesis aims to investigate whether the IEKF demonstrates improved performance over the MEKF in 3D pose estimation using inertial measurement units (IMUs).

Specifically, the main research question of this thesis is: How does the estimation accuracy of the invariant EKF compare to the multiplicative EKF in the context of pose estimation? In order to gain insight into this, the investigation focuses on three main questions. Firstly, what are the advantages of utilizing a left-invariant EKF (LIEKF) over an MEKF when dealing with a left-invariant measurement model, and similarly, what are the benefits of employing a rightinvariant IEKF (RIEKF) over an MEKF when dealing with a right-invariant measurement model? Secondly, how does the IMU sensor noise magnitude affect the converging performance
of the filters differently? Thirdly, How does the sensor noise magnitude of the external measurements affect the converging performance of the filters differently?

Additionally to the distinction between the left- and right-IEKF, a similar distinction is made for the MEKF. This thesis distinguishes between an MEKF with orientation deviation states resolved in the body frame (MEKF-b) and an MEKF with orientation deviation states resolved in navigation frame (MEKF-n). This distinction is made since it allows for a more natural comparison between the IEKF and MEKF.

To conduct the evaluation, extensive simulations are performed, allowing for controlled variations in these parameters. The simulation results provide insights into the comparative performance of the IEKF and multiplicative EKF under different conditions, shedding light on their strengths and limitations in 3D pose estimation with IMUs.

It was found that the IEKF and MEKF show very comparable results in a large amount of the applications. The state-independent error dynamics have been shown to be beneficial in situations where the initial information of the state of the system is uncertain. Furthermore, the IEKF has been shown to be beneficial in certain edge cases. Firstly, the IEKF shows to be less sensitive to small process noise covariance matrices Q. Secondly, once the gyroscopic noise becomes very large, the RIEKF showed higher estimation accuracy over the MEKF-n. The LIEKF did also show a marginal improvement in estimation accuracy over the MEKF-b.
Finally, it was found in this thesis there are two ways that the external measurement noise influenced the comparison of the estimation accuracy between the IEKF and MEKF. The MEKF-n showed to be sensitive to a low covariance measurement matrix R and additionally, the MEKF-b and MEKF-n both seemed to be marginally more affected by higher external measurement noise than the LIEKF and RIEKF, respectively.

In conclusion, this thesis provides a comprehensive evaluation of the IEKF and MEKF in 3D pose estimation with IMUs. While the IEKF and MEKF exhibit comparable performance in many cases, the IEKF’s state-independent error dynamics and its advantages in certain scenarios highlight its potential superiority over the MEKF. These findings contribute to the understanding of nonlinear state estimation on matrix Lie groups and offer valuable insights for selecting the appropriate filter for specific applications. ... Read more