In the last decades, the Synthetic Aperture Radar (SAR) systems have become among the most reliable remote sensing instruments for monitoring the Earth's surface, due to their ability of large-scale analysis with ease. Precise results can be achieved due to the continuous improvement of sensors' capabilities and the development of coherent processing techniques, like SAR Interferometry and Tomography. A popular approach from this topic is based on the detection of stable scatterering mechanisms. In this sense, two detection algorithms, CAESAR-D and SqueeSAR-D have been recently developed by adapting efficient interferometric phase filtering techniques. The relative performances of those detectors have been evaluated and compared under multiple hypothesis. This work represents a continuation of those studies, by analyzing the capabilities of the two algorithms in correspondence with the sensors' characteristics, this representing the main contribution of the paper. In this sense, two datasets of the same test area, acquired by different SAR sensors, Cosmo-SkyMed and Sentinel-1, are exploited.
With unmanned aerial systems (UAS) becoming more commonplace, the threat they pose also increases. Therefore, Counter-UAS (C-UAS) systems become increasingly important to protect critical infrastructure and/or public events. Because UAS are difficult targets to detect, multi-sensor systems are a preferred means of detection. In order to achieve a successful fusion of the involved sensors' data, the system needs to be calibrated. For non-stationary applications the C-UAS system has to be set up each time it is deployed and needs to be extrinsically calibrated each time as well. The process of setting up and calibrating is typically conducted by the operator of the system, who often are not experts in the sensors themselves. Therefore, we propose an automated calibration process using a cooperative UAS to estimate each sensors' biases in azimuth and elevation. Additionally, different mathematical measures are used to automatically interpret the success of the calibration and provide additional information to the operator. The method is applied and evaluated on real-world data, and its accuracy and robustness are assessed.
Datasets are a crucial element in the development of perception algorithms. They relate sensor measurement data to annotated reference information and allow for the deduction of sensor and object characteristics. In autonomous driving, the reference data commonly consist of semantic image segmentation, point-wise associations, or bounding box annotations. The dataset proposed in this work, however, aims to dig deeper into the evaluation of measurement principles and provides scanned 3D models of all vehicles together with a pose and continuous kinematics reference obtained by RTK-GNSS. Combined, the state of the complete dynamic surrounding of the sensor vehicle is known for any point in time. Subsequent reference formats can be easily computed in user-defined granularity. This dataset involves single-object and multi-object recordings with seven target vehicles. In particular, measurement effects such as occlusion, as well as reflections, can be evaluated, as the normals of the shape of the target vehicles are known. We describe the dataset, discuss the technical background of its development, and briefly present exemplary evaluations.
We present a gate-based, digitized adiabatic (QAOA-style) realization of the Bayesian filtering update on discrete grids. Our method encodes the likelihood into a diagonal cost Hamiltonian and implements the anneal via a symmetric Trotter sequence of mixer and cost blocks, avoiding state-conditioned gate constructions and any variational parameter training. We show theoretically that the measurement statistics exhibit the same effective inverse temperature as in the continuous anneal up to O(T-3 /p(2)) Trotter error, and that a single energy-scale parameter alpha(applied only to the cost) restores the target posterior variance after a one-shot calibration. On Gaussian test cases, the calibrated digitized anneal closely matches the analytical posterior while maintaining shallow, structured circuits that scale naturally with the register size. These results position digitized adiabatic filtering as a practical alternative to both analog annealing and state-conditioned gate approaches in quantum data fusion.
In this work, we establish the consistency and asymptotic efficiency of an online formulation of the maximum likelihood estimator for independent, non-identically distributed random variables. The proposed online approach has a computational cost of O(1) per measurement, offering a significant improvement over the O(n) cost required by the traditional formulation after the n-th observation. We demonstrate the applicability of the developed theory to the Direct Position Determination geolocation problem. Simulation results confirm that, under practically relevant conditions, both the online and traditional formulations achieve near-optimal estimation performance, while the online algorithm shows clear advantages in terms of computation time.
Statistical methods to detect signals of after-market adverse drug reactions from Spontaneous Reporting System (SRS) are a subject of active research in pharmacovigilance. SRS data are considered as the cornerstone in signal detection, but complementary sources of data, including social media, have also been considered. This paper proposes a Beta-Binomial distribution-based Bayesian signal detection method for fusing SRS data with social media data, and applies it to combining records from the FDA Adverse Event (AE) Reporting System and Twitter. We show that, with careful processing, particularly with respect to bias potentially introduced by the presence of high counts of drug-AEs ('masking'), fusing the multi-source data can improve performance relative to using either source alone.
Joint target tracking and sensor scheduling includes resource optimisation and gathering the most informative data for purposes such as search and rescue, fire detection and surveillance tasks. For such real-time tasks, the limited access to initial tracking data can challenge the effectiveness of traditional machine learning methods, thereby motivating the development of active sensing strategies. This paper addresses such problems and formulates the joint target tracking and sensor scheduling problems within a Bayesian optimisation framework. The key question that this framework answers is: where to position the sensors in order to accurately track an object. In the considered case study, the sensors are mobile and represented by uncrewed aerial vehicles (UAVs). The active sensing of the environment is based on uncertainty-guided sampling thanks to a Gaussian process representation. The main novelty lies in the formulation of the sensor scheduling and tracking within a Bayesian optimisation setting. Under this framework, a detailed comparison of different acquisition functions is carried out, to identify the most suitable solutions for an active sensing problem. Results with respect to accuracy and computational time are reported.
Maritime safety remains vulnerable to GNSS spoofing, limited detection ranges, and environmental uncertainties, even with the use of positioning systems such as AIS and shore-based radar. To address these challenges, previous work introduced SeaSentry, a passive shore-based sensor network designed to detect and monitor vessels. The placement of sensors within the network significantly impacts the accuracy of vessel localization and tracking, as both the number and spatial configuration of sensors determine positioning performance. This study proposes a grid-based search for optimizing sensor placement in scan-based vessel localization. A network of fixed sensors is analyzed to identify the optimal positions for additional sensors, using the Cramer-Rao Lower Bound (CRLB) to quantify the minimum achievable localization error. Experiments conducted in the Neuburgweier region of Germany demonstrate the potential of the proposed grid-based search approach to establish a robust optimization technique for practical deployment. The findings also underscore the importance of strategic sensor placement and validate the approach through empirical results.
Effective maritime border surveillance is crucial. Challenges we face include irregular migration, smuggling, oil spills and the need for rapid search and rescue. Various sensing technologies, including AIS, SAR, optical and infrared sensors, as well as UAV-mounted sensors, clearly enhance maritime awareness. However, integrating their diverse outputs remains complex. Feature-level multi-modal sensor fusion is a well-known methodology for robust detection and behavior analysis. However, most research relies on simulations or isolated sensors, which limits practical insights. This study presents a controlled real-world experiment combining synchronized data from coastal ground sensors and UAVmounted visual and infrared sensors. The recorded dataset enables the evaluation of feature-level fusion in authentic conditions. We enhance existing fusion frameworks with additional modules and assess them using operational metrics. This study contributes to our understanding of the efficacy of multi-modal fusion in complex maritime environments, while also highlighting the significant challenges involved in transitioning from simulations to controlled real-world sensor data.
In multi-sensor multi-target tracking, the task of track-to-track-association (T2TA) is to identify tracks from different sensors that belong to the same object. For noisy environments with massive raw measurements, two-stage clustering solutions have been suggested where initial clusters are repartitioned in a second step. To explore the potential of clustering for T2TA, we adapt and enhance this concept by designing a novel multi-stage T2TA framework where constraint-violating clusters are repartitioned and singletons or outliers are re-evaluated based on a likelihood function. We present experiments with up to 20 sensors and various detection probabilities in complex scenarios including overlapping target tracks. Our approach achieves promising accuracy and outperforms other clustering methods in both sparse and dense environments. In comparison to stochastic optimization (SO), a sampling-based technique, we do not achieve the same accuracy in sparse environments, but can still take advantage of clustering by integrating SO as repartitioning method. Even without specification of a fixed gating threshold, this hybrid version is able to maintain comparable accuracy while reducing the computationally intense association part.
The Fokker-Planck propagator is derived for prediction on cylindric manifolds. We exploit the low-rank tensor decomposition technique that is already being used in the Euclidean domain. With only a small change to the finite difference matrix, we can readily apply it to certain manifolds such as the cylinder. Our application example is estimating the angular position and velocity of a rotating shaft. This state estimation problem may seem linear at first glance, but since the underlying state space is nonlinear due to the periodicity of the angular coordinate, it is an inherently nonlinear estimation problem.
A STAP algorithm is described, that reduces the signal distortions which are introduced by a standard STAP filter. The reduction is accomplished by introducing additional linear constraints on the space-time filter coefficients. A proper selection of the linear constraints leads to a STAP filter with a linear-phase characteristic (i.e. a filter impulse response with constant group delay). Thereby, the STAP algorithm reduces the signal distortions at the output of the STAP filter. A least-meansquare adaptive filter version of the STAP filter with linear-phase constraints is proposed. The feasibility of the proposed adaptive STAP algorithm with linear-phase characteristic is analyzed based on Monte-Carlo-Simulations.
Estimating causal mechanisms from multi-domain data presents significant challenges, particularly when the independent and identically distributed (i.i.d.) assumption for most existing causal discovery methods no longer hold. To address this, the Common and Individual Causal Mechanism Estimation (CICME) approach introduces a strategy for identifying stable variables with domain-invariant causal mechanisms and recovering domain-specific causal mechanisms in each domain. In this work, we propose an improvement by introducing an intermediate fine-tuning step, which effectively improves the accuracy of stable variable detection and the resulting causal structures. We consider both linear and nonlinear data in the evaluation scenarios. Additionally, we assess the practical value of the resulting models through two manufacturing tasks: quality prediction and root cause analysis (RCA). Based on the results, we offer actionable guidelines for practitioners applying machine learning models in multi-domain settings.
This study investigates fall detection using only Ultra-Wideband (UWB) distance data. To generate reliable training labels, a YOLOv8-based vision model was employed to automatically detect fall events from video, which were then aligned with the UWB measurements. After preprocessing steps including imputation, noise reduction, interpolation, and scaling, class imbalance in the dataset was mitigated through balancing techniques. The UWB sequences were then classified by a hybrid CNN-GRU model. Tested on subjects not included in the training set, the approach achieved 94% accuracy and 93% recall, demonstrating strong cross-subject generalization. The results highlight the potential of UWB-based systems, supported by vision-assisted labeling, as non-intrusive and privacy-preserving solutions for healthcare and safety applications.
We analyze the use of Long Short-Term Memory (LSTM) networks and Gaussian LSTM networks (G-LSTMs) for object localization and tracking based on bistatic sonar measurements in scenarios involving moving receivers, missing detections, and potentially biased Gaussian measurements. We formally derive the Cramer-Rao Lower Bound (CRLB) for the situation including incorporating past information for noisy non-linear models. Performance is analyzed on decaying coordinated turn tracks and found to be close to the CRLB and the performance of the Extended Kalman Filter (EKF), despite having no explicit prior knowledge of the movement or measurement models. The Gaussian and non-Gaussian LSTMs exhibit robustness against missing detections and measurement bias and are able to quantify such biases (if present). This Gaussian architecture shows promise, as performance is slightly worse with their non-Gaussian counterparts but offers a covariance estimate and only exhibits a slight underconfidence in predictions. The Gaussian neural networks considered here show the ability to capture environmental uncertainty as given by the dynamic probability of detection. Inference times are small enough to allow for real-time tracking.
Accurate Multi-Target Tracking (MTT) in cluttered environments remains a significant challenge due to the coupled nature of state estimation and data association. Since errors in state estimation and data association reinforce one another, reliable MTT requires interpretable methods that expose how uncertainties propagate across the two tasks, enabling robust and trustworthy tracking. In this work, we investigate the robustness of a previously proposed ensemble-based tracking system for maneuvering target tracking, called Ensemble of KalmanNets. By drawing inspiration from the Interacting Multiple Models filter and Bayesian recursive estimation, the proposed approach preserves the interpretability of the filtering process. The system combines neural networks integration in the filtering process, with the simple Global Nearest Neighbor association method, enabling transparent optimization of measurement-to-target assignments through Mahalanobis distances. We demonstrate how improved state estimation reinforces consistent associations and, conversely, how robust association improves tracking accuracy. Extensive evaluations show that our system maintains high performance in challenging cluttered scenarios while keeping the decision-making process fully explainable, providing a practical solution for real-time and interpretable MTT applications.
Track evaluation in sensor data fusion typically relies on assignment-based ground truth association to compute metrics. Although widely adopted, this approach overlooks temporal continuity and can produce misleading performance assessments. This paper highlights the differences and presents a systematic comparison of assignment-based and time-based association in metric computation, with a focus on the target completeness metric, and offers guidance for researchers and practitioners to achieve more reproducible and comparable results. An evaluation is also formed from the perspective of different output frequency of the trackers and its impact on the final metric score. The results show that inconsistencies created using the assignment-based approach can easily reach beyond 20 %. Explanation of when this could happen, how to decide if it happens in a given situation, and the strategy of how to fix this, is provided.
Cooperative airborne sensor deployment enhances situational awareness by combining complementary sensor carrying platforms to improve target detection, localization, and classification, while overcoming individual limitations and optimizing resource use. To enable automated coordination of heterogeneous aircraft and sensors, sensor fusion and scheduling are essential. Deployment and fusion are interdependent, as sensor placement influences the quality of fused data, while fused data can guide the allocation of additional sensor resources for improved coverage and accuracy. This work focuses on using Ground Moving Target Indicator (GMTI), Electronic Support Measures (ESM) instrumentation and Electro-Optical/Infrared (EO/IR) sensors with a scheduling logic that allocates resources dynamically based on fusion results and situational needs. Measurements are fused via a Joint Probabilistic Data Association (JPDA) algorithm with a Bayesian classifier, supported by sensor performance models estimating measurement quality. The integrated approach is evaluated in a simulated wide-area Intelligence, Surveillance, and Reconnaissance (ISR) mission.
This study examines eye-tracking sensor data from 121 participants interacting with six websites to evaluate the predictive value of different eye movement features for perceived usability and user experience (UX). In total, over 36 hours of eye-tracking data, recorded at 250 Hz and totaling 32 million gaze points, were transformed into second-order eye movements events and then into third- and fourth-order machine learning features. Usability and UX ratings, collected through short versions of the User Experience Questionnaire (UEQ) and AttrakDiff questionnaires, served as labels for training six machine learning models. Results show that spatio-temporal metrics, particularly saccade sequences and Area of Interest (AOI) transitions, are the most predictive. Models trained on fourth-order features consistently outperform those based only on aggregated third-order metrics, highlighting the importance of preserving scanpath structure and sequential gaze behavior for predictive performance in eye-tracking-based Human-Computer Interaction (HCI) research.
Data fusion techniques combine information about an object that is observed by multiple sources. While the fusion of raw measurements provides the best results, the radio link is often limited and restricts the sources to transmitting condensed track data. In this paper, we consider the case that the sources utilize Interacting Multiple Model filters to track the object and the radio link is capable of the transmission of the mixture data. We treat this topic from an application perspective and assume a large number of sources and a fusion center without knowledge about internal filter parameterization. Our aim is to fuse all Interacting Multiple Model modes of all sources for a preferably high estimation quality. To manage the potentially exponential number of mixture components, we perform a modewise fusion and propose a tree structure that enables the reuse of previously processed data. We compare this approach with a central measurement fusion and a moment-matched fusion employing a Monte-Carlo simulation to evaluate the estimation quality. Especially during changes in the motion behavior of the object, a moderate improvement in the estimation quality can be observed when compared to the moment-matched fusion.
