ABSTRACT. Marginalization enables the particle filter to be applied to non-trivial problems by invoking the Kalman filter to estimate a larger part of the state vector. The marginalized (a.k.a. Rao-Blackwellization) particle filter (MPF) has found many use cases in tracking and navigation applications. These are characterized of having position and its derivatives as states. Here, we take a closer look at the MPF for the constant velocity motion model, which well represents the basic properties of most motion models used in this context. We investigate in particular how the Kalman Filter (KF)-part depends on how the continuous time state noise is sampled in the discrete time model. Rather unexpectedly, the KF degenerates for many of the most common sampling approaches, in the sense that the covariance approaches $0$. We also show that for those same sampling approaches there is no performance increase by switching from a Particle filter to a MPF.

ABSTRACT. In tracking applications, multiple radars used to observe target motion usually work asynchronously due to different sampling rates and initial sampling time instants, and the fusion time instant at the fusion center can be designated arbitrarily. In this paper, the estimation fusion problem for target tracking with asynchronous multi-rate multi-radar measurements is investigated. Two asynchronous fusion algorithms are proposed, i.e., batch time-aligned asynchronous fusion with unbiased converted measurements and sequential linear minimum mean square error (LMMSE) asynchronous fusion with converted measurements. The batch time-aligned asynchronous fusion algorithm is suboptimal because of the correlation between measurement error covariance and measurement itself. The sequential LMMSE asynchronous fusion algorithm is theoretically optimal in the sense of minimizing the mean square error within the set of all linear estimators. Numerical examples are provided to demonstrate the effectiveness of the proposed two asynchronous fusion algorithms.

ABSTRACT. The paper considers the solution to the state estimation problem of nonlinear dynamic stochastic systems by the particle filters. It focuses on the marginal particle filter algorithms which generate samples directly in the marginal space for the recent state. Their standard implementation calculates the sample weights by combining the samples from two consecutive time instants in the transition and proposal density function evaluations. This results in computational complexity quadratic in sample size. The paper proposes an efficient implementation of the marginal particle filter for which a functional tensor decomposition of the transition and proposal densities is calculated. The computational complexity of the proposed implementation is linear in sample size and the decomposition rank can be used to achieve a trade-off between accuracy and computational costs. The balance between the complexity and the estimate quality can be tuned by selecting the rank of the decomposition. The proposed implementation is demonstrated using two numerical examples with a univariate non-stationary growth model and terrain-aided navigation scenario.

ABSTRACT. Stochastic particle flow filters are driven by diffusion processes. Recently it has been discovered that a stochastic particle flow filter exists for any given diffusion matrix. Therefore, the diffusion matrix effectively acts as a design parameter. A variety of diffusion matrices have been proposed in the literature. One natural question is what are the differences among various choices and how to select the diffusion matrix to achieve optimal filtering performance. Particle flow filters are described by continuous-time stochastic differential equations and realized through numerical discretization whose accuracy is affected by both the stiffness of the differential equations and the value of the diffusion matrix. In this paper, we examine the role of the diffusion matrix in stiffness mitigation of stochastic particle flows and its trade-off with numerical accuracy. Objective functions are proposed to balance several conflicting factors of stiffness mitigation and numerical accuracy. Analytic solutions of optimal diffusion matrices are derived for stiffness mitigation of stochastic particle flows. Results are illustrated using numerical examples.

ABSTRACT. We propose a method for deterministic sampling of arbitrary continuous angular probability density functions. With deterministic sampling, good estimation results can be obtained with a much smaller number of samples than with the commonly used random sampling. The Unscented Kalman Filter also uses deterministic sampling, but takes a very small number of samples. Our method can draw an arbitrary number of deterministic samples, improving the quality of state estimation. Conformity between the continuous density function (reference) and the Dirac mixture density, i.e., sample locations (approximation), is established by minimizing the distance of the cumulatives of dozens of univariate projections. In other words, we compare  density functions in Radon space.

ABSTRACT. This paper defines a working ontology to represent entities across all fusion levels, as well as for corresponding entity states and state estimates at all fusion levels. The proposed ontology involves a common representational scheme for simple and complex structures at all fusion levels, thereby facilitating inference within and among data fusion levels. Formalisms are defined for consistent representation of simple and complex entities, entity states, and state estimates at all fusion levels. Events are construed as state change relationships, and therefore as entities subject to estimation. By reifying relations and relationships (and treating attributes as attributes as 1-dimensional versions of these), we can apply the whole battery of level 1 fusion machinery to level 2 and 3 problems. This has advantages of design reuse and synergy across the levels. We illustrate such belief propagation in a simple Electronic Combat situation.

ABSTRACT. Drone applications in military and civil fields continue to be of great interest, but all of them require a critical and essential maneuver, landing. Although the problem of autonomous landing is not new, the continuous technological advances in sensorization, navigation systems, or data processing among others, encourage the emergence of new strategies continuously. A widespread approach in the literature for multirotor vehicles are based on computer vision. However, from the simplest system with monocular vision to the most complex with multiple sensors, it can be improved if its behavior is studied. This paper evaluates the global position estimation accuracy of a landing pad provided by an autonomous landing system using information from the aircraft navigation system, the vehicle attitude, and a gimbal-adjustable monocular vision system. The results of this study show that bias correction over a polar space improves a set of quality metrics that evaluate the performance of the landing phase.

ABSTRACT. In recent years the idea of fusing diverse type of information has often been employed to solve various Deep Learning tasks. Whether these regard an NLP problem or a Machine Vision one, the concept of using more input data of the same type has been the basis of automated studies. Considering NLP problems, approaches involving the merging of different word embeddings have already been tried, managing to make improvements to the most common benchmarks. Here we want to explore the combination not only of different types of input together, but also different data modalities. This is done by fusing two popular word embeddings together, mainly ELMo and BERT, with other inputs that embed a visual description of the analysed text. Doing so, different modalities -textual and visualare both employed to solve a textual problem, a concreteness task. Multimodal feature fusion is here explored through several techniques: input redundancy, concatenation, average, dimensionality reduction and augmentation. By combining these techniques it is possible to generate different vector representations: the goal is to understand which feature fusion techniques allow to obtain more accurate embeddings.

ABSTRACT. Robotic systems operating in the real world would benefit from a clear semantic model to understand their interactions with the real world. Such semantics are typically captured in an ontology. Unfortunately, the underlying model of existing ontologies requires many work-arounds before it can be used to capture general knowledge about objects and interactions in the real physical world. To remove such work-arounds, this article adopts the richer hypergraph model. It is used to develop an ontology, which is further implemented as the knowledge base for an actual robotic system performing search operations. Actual information extracted from the robot's sensors is used to update its knowledge base logically and sensibly.

ABSTRACT. An innovative framework, based on the symmetric Sugeno integral, is proposed in this paper to calculate confidence aggregation. The symmetric Sugeno integral implements nonadditive (fuzzy) integration and is shown to have the desired and unique properties to handle contradictory evidence, support commutivity (solution invariance under the ordering of aggregation sequences) and demonstrate idempotence (identity operations). One size does not fit all in the proposed solution: Each community of domain-specific analysts will use their own, local mapping to a shared scale which is valid across the entire analyst enterprise. This paper contributes the first known approach to the long-standing confidence aggregation problem that is undertaken purely in the ordinal domain, without empirical projection onto the numerical domain. The paper also describes the first known application of the (symmetric) Sugeno integral to intelligence evidence handling. Worked examples exercise the proposed solution to an unaddressed but sought capability in analytical tradecraft, which recognises the diversity of evidence set contributions to different intelligence tasks.

ABSTRACT. Within an ultra-tightly (or deeply) integrated global navigation satellite system (GNSS) and inertial navigation system (INS) GNSS/INS, GNSS signal correlation delivers correlator values as input to the integration filter. On the other side, the integration filter controls the correlation process by determining the numerically controlled oscillator (NCO) values. As GNSS signal correlation is a computational trivial but a time-consuming process, we propose for R&D in this area an alternative approach to first generate for each GNSS signal multi-correlator values and store them for the later GNSS/INS filter development work.Once the filter runs, it interpolates from the multi-correlator values the actual needed correlation values. The multi-correlator values thus act like a data compression for the GNSS signals.This paper discusses the mathematical framework for this data compression, which is loosely described as a sufficient statistic.This statistic consists of the correlation values themselves plus the NCO values that have been used during the correlation process.The generation and interpolation process will be describedin this contribution with all mathematical details, as well as interpolation limits in code phase and Doppler direction. Finally, this approach is validated by comparison of global positioning system (GPS) C/A code pseudorange and carrier phase datafrom direct tracking to results originate from a MATLAB-based receiver which uses the multi-correlator values as sufficient statistics.

ABSTRACT. The aim of this article is to present a novel parallelization method for temporal Gaussian process (GP) regression problems. The method allows for solving GP regression problems in logarithmic O(log(N))time, where N stands for the number of observations and test points. Our approach uses the state-space representation of GPs which, in its original form, allows for linear O(N) time GP regression by leveraging Kalman filtering and smoothing methods. By using a recently proposed parallelization method for Bayesian filters and smoothers, we are able to reduce the linear computational complexity of the temporal GP regression problems into logarithmic span complexity. This ensures logarithmic time complexity when parallel hardware such as a graphics processing unit (GPU) are employed. We experimentally show the computational benefits of our approach on simulated and real datasets via our open-source implementation leveraging the GPflow framework.

ABSTRACT. Abstract—Aiming at the problem of multi-target tracking under the condition of noises statistics mismatch, an adaptive δ-GLMB filter based on VB approximation is proposed. The Normal-inverse Wishart distribution is used to model the state one-step prediction and prediction error covariance matrix, and the joint distribution of mean and covariance matrix of measurement noise, and the latent variables are described as Gamma distribution. In this paper, the filter density of single target is expressed as the mixture of Normal inverse Wishart inverse Wishart Gamma Gamma (NNIWNIWGG), and its NNIWNIWGG mixture implementation under linear Gaussian condition is given. According to the minimization of Kullback-Leibler divergence, the approximate solution of marginal likelihood function is obtained. Simulation results show that the proposed adaptive δ-GLMB filter has high tracking accuracy in the case of noises statistics mismatch.

ABSTRACT. A hybrid analytical-machine learning (ML) framework for improved low Earth orbit (LEO) satellite orbit prediction is developed. The framework assumes the following three stages. (i) LEO satellite first pass: A terrestrial receiver with knowledge of its position produces carrier phase measurements from received LEO satellite signals, enabling it to estimate the time of arrival. The LEO satellite’s states are initialized with simplified general perturbations 4 (SGP4)-propagated two-line element (TLE) data, and are subsequently estimated via an extended Kalman filter (EKF) during the period of satellite visibility. (ii) LEO satellite not in view: a nonlinear autoregressive with exogenous inputs (NARX) neural network is trained on the estimated ephemeris and is used to propagate the LEO satellite orbit for the period where the satellite is not in view. (iii) LEO satellite second pass: a terrestrial receiver with no knowledge of its position uses the ML-predicted LEO ephemeris along with its carrier phase measurements from received LEO signals to estimate its own position via an EKF. Experimental results with with signals from an Orbcomm satellite are presented to demonstrate the efficacy of the proposed framework. It is shown that during the satellite’s second pass, the ML-predicted ephemeris error is reduced by nearly 90% from that of an SGP4 propagation. In addition, it is shown that if the receiver was to use the SGP4-predicted satellite ephemeris to localize itself, the EKF’s initial position error of 2.2 km increases to 6.7 km, while the proposed framework reduces the position error to 448 m.

ABSTRACT. Cross-correlation generally exists between local estimation errors in distributed estimation fusion but is hard to know exactly. In some cases, the cross-correlation is partially known (e.g., only correlation level is known approximately), or known within an interval of correlation intensity. Utilizing correlation information benefits estimation. To use correlation knowledge, one way is to model correlation as the generalized Pearson's correlation coefficient times the matrix product of local MSE matrices. The correlation coefficient is used to measure the linear relation intensity between two random vectors, and is assumed to be randomly located in the provided interval, e.g., uniformly distributed. Based on the way the cross-covariance matrix is modeled, the assumption about correlation coefficient, and by applying the best linear unbiased estimation (BLUE) fuser, an estimation fusion algorithm, called expected BLUE fuser (EBF), is presented. Comparing with similar algorithms, its comparable or better performance demonstrates the effectiveness of the proposed algorithm. Considering the fusion results under various given correlation intervals, we observe that strong correlation intensity benefits fusion performance improvement, and point out that EBF gets good fusion results when the given correlation interval cover the true correlation coefficient closely.

ABSTRACT. The use of Unmanned Aerial Vehicles (UAV) has increased in recent years. Increased density of air traffic as well as the autonomy of the vehicles involved, demand robust safety of traffic operations in terms of dependable decision making for flight operations. Since future traffic management services (U-space) will focus on registration, identification, approval to fly, etc., and cooperative traffic avoidance such as FLARM requires that other parties be equipped as well, future UAVs should be able to robustly detect uncooperative parties and avoid mid-air collisions in airspace. To ensure the highest robustness and to increase sensitivity and accuracy, a combination of several sensors systems by multi-sensor data fusion techniques is highly recommended. This paper formulates a novel multi sensor data fusion algorithm, that is a joint approach of multiple hypothesis tracking algorithm and particle filtering. The union of these two algorithms combines the strength of the multiple hypothesis tracking for data association with the robustness of the particle filter to estimate the position of the tracked objects. This joint approach has been validated with the use of simulated data.

ABSTRACT. Finding the most probable posterior hypotheses is a core task in hypothesis-oriented multiple hypothesis tracking (HO-MHT), and also in related tracking methods such as the Poisson multi-Bernoulli mixture (PMBM) filter. The traditional approach is to find the M best new hypotheses for each parent hypothesis by means of Murty’s algorithm. In this paper we instead present an algorithm for finding the M best hypotheses ranging over all parent hypotheses. The algorithm is developed in the more general context of cluster management, where the goal is to merge several parent clusters, and to find the M best posterior hypotheses in any such supercluster.

ABSTRACT. Though significant progress has been achieved in the mathematical theory of multi-target tracking and its application in numerous surveillance domains, robust solutions are not always achieved in practice. This paper offers design suggestions for improved performance with a primary focus on multiple-hypothesis tracking based methods.

ABSTRACT. Incorporating prior environmental information to the diversely applied field of target tracking is becoming more essential in order to leverage maximum sensing performance. Urban scenarios in particular present numerous large obstacles that will block the field of view of most sensors along with many potential sources of measurement clutter. Handling these complications appropriately is imperative to ensure improved target tracking performance. This paper presents a computationally efficient method of integrating visibility information within the Bernoulli particle filter. Through estimation of target visibility with ray casting, the probability of target detection, birth density and spatial target density are modified. Numerical results demonstrate significantly more gradual degradation in target state estimation performance and improved estimation of target existence in the occlusion situation. Faster tracking recovery when emerging from occluded regions is also demonstrated in the results.

ABSTRACT. This paper considers multitarget tracking under out-of-sequence measurements (OOSMs), i.e. when the measurements processed by the tracker might be out of order. In order to fully exploit information provided by the sensors, OOSMs should be re-utilized rather than being simply discarded so as to improve tracking performance. To this end, this paper proposes a message passing (MP) multitarget tracking algorithm under OOSMs, where MP is adopted to perform efficient association between target and (in-sequence and out-of-sequence) measurements. Simulation experiments show that, compared to simply discarding OOSMs, the accuracy in terms of target number and state estimates can be greatly enhanced by incorporating OOSMs, thus demonstrating the effectiveness of the proposed approach.

ABSTRACT. In this paper, an extended Kalman filter (EKF) framework, called uncertainty aware EKF (UA-EKF), is developed by utilizing contextual knowledge to improve the vehicle tracking accuracy for autonomous vehicles. The proposed framework can learn and estimate the uncertainty associated with the measurements provided by a LiDAR-based vehicle detection algorithm. The UA-EKF has two major parts: one has the ability to estimate the state-dependent measurement noise’s statistics for LiDAR object detections, and the other is to create multiple-hypothesis measurements based on the detected vehicle’s heading. The measurement uncertainties are learned based on the EKFNet, which is an algorithm that can learn the system noise covariance from measurement data. Both the learned noise statistics and multiple-hypothesis estimators are used to compensate the physical limitations of the LiDAR measurements. A detailed analysis of the measurement uncertainty and the methods to improve tracking performance during filtering are provided for the UA-EKF. The obtained results by using the nuScenes datasets show that estimating the measurement uncertainty is an efficient solution for tracking the vehicle based on LiDAR detections.

ABSTRACT. We present a new method to combine evidential top- view grid maps estimated based on heterogeneous sensor sources. Demster’s combination rule that is usually applied in this context performs poorly with highly conflicting inputs. Therefore, we use more advanced evidential reasoning (ER) techniques and improve the conflict resolution by modeling the reliability of the evidence sources. We propose a data-driven reliability estimation to optimize the fusion quality based on the Kitti-360 dataset. We apply the proposed method to the fusion of light detection and ranging (LiDAR) and Stereo Camera data and evaluate the results qualitatively and quantitatively.

ABSTRACT. To achieve object detection on low-computing devices such as embedded and mobile devices, we propose a new method of the non-deep network for fast vehicle detection. In our method, we use color transformation to address the problem of insufficient training data. To achieve effective object detection of different sizes, we introduce a non-deep network with a parallel double-stream design. The upper stream adopts the downsampling block which contains a 3*3 kernel size convolution layer to extract the feature of small objects. The lower stream uses the downsampling block which contains a 5*5 kernel size convolution layer to realize the detection of large targets. Finally, these extracted features of different receptive fields are fused in the fusion block. We use the one-level output feature from backbone for detection to improve model efficiency. The experimental results show that our network runs real-time on Jetson TX2, and achieves 30.46% mAP on COCO and 77.2% mAP on UA-DETRAC. Our detector is more accurate than YOLO-Fastest and faster than YOLOv4-tiny.

ABSTRACT. Autonomous vehicles depend on an accurate perception of their surroundings. For this purpose, different approaches are used to detect traffic participants such as cars, cyclists, and pedestrians, as well as static objects. A commonly used method is object detection and classification in camera images. However, due to the limited field of view of camera images, detecting in the entire environment of the ego-vehicle is an additional challenge. Some solutions include the use of catadioptric cameras or clustered surround view camera systems that require a large installation height. In multi-camera setups, an additional step is required to merge objects from overlapping areas between cameras. As an alternative to these systems, we present a real-time capable image stitching method to improve the horizontal field of view for object detection in autonomous driving. To do this, we use a spherical camera model and determine the overlapping area of the neighboring images based on the calibration. Furthermore, lidar measurements are used to improve image alignment. Finally, seam carving is applied to optimize the transition between the images. We tested our approach on a modular redundant sensor platform and on the publicly available nuScenes dataset. In addition to qualitative results, we evaluated the stitched images using an object detection network. Moreover, the real-time capability of our image stitching method is shown in a runtime analysis.

ABSTRACT. Much research has been done on the detection and tracking of paved, preferably marked roads. Less work is available on the detection of dirt roads. The challenge is to provide a framework to track both paved roads and dirt roads. In this paper, we are addressing the problem of developing measurement approaches working for both kinds of roads likewise. For that we fuse LiDAR with vision: First, we present indirect measurements from a static environment model populated with LiDAR data, as well as a new approach for LiDAR measurements from a segmented point cloud. Second, we investigate different image color modes to improve the effectiveness of locating dirt road boundaries using local oriented edge detection. We demonstrate the robustness of our measurements on difficult roads by showing qualitative results from our autonomous vehicles.

ABSTRACT. The state estimation of repetitive processes with periodically repeated trajectories can be interpreted as the dual task of iterative learning control design. While the latter has been widely investigated over the last two decades, only few approaches exist for the design of iterative learning observers. However, the exploitation of the knowledge about periodically repeated trajectories, which occur among others in pick and place tasks in robotics as well as in charging and discharging of batteries, offers the opportunity to enhance the estimation accuracy from one execution of the control task to the next. In this paper, we generalize a linear stochastic approach for iterative learning state estimation, inspired by the Kalman filter in terms of a minimization of the estimation error covariance, to the class of models with bounded parameter uncertainty and to nonlinear ones that can be represented by means of quasi-linear discrete-time state-space representations. To solve this task, a novel combination of set-valued ellipsoidal state enclosure techniques with the aforementioned stochastic iterative learning state estimator is presented and visualized for a quasi-linear model of the charging/discharging dynamics of Lithium-Ion batteries.

ABSTRACT. The reapproximation of discrete probability densities is a common task in sample- based filters such as the particle filter. It can be viewed as the approximation of a given Dirac mixture density with another one, typically with less samples. In this paper, the Wasserstein distance is established as a suitable measure to compare two Dirac mixtures. The resulting minimization problem is also known as location-allocation or facility location problem and cannot be solved in polynomial time. Therefore, the well-known sliced Wasserstein distance is introduced as a replacement and its ties to the \ac{pcd} are shown. An iterative algorithm is proposed to minimize the sliced Wasserstein distance between the given distribution and approximation.

ABSTRACT. This paper deals with the state estimation of non-linear stochastic dynamic systems with an emphasis on a probability density function approximation used by point-mass filters. Approximation error of the standard point-mass density is analysed and quantified, and a novel point-mass density approximation with inherent approximation error minimisation is developed. The properties of the proposed point-mass are theoretically analysed and numerically illustrated.

ABSTRACT. In this paper we present a hybrid neural network augmented physics-based modeling (APBM) framework for Bayesian nonlinear latent space estimation. The proposed APBM strategy allows for model adaptation when new operation conditions come into play or the physics-based model is insufficient (or incomplete) to properly describe the latent phenomenon. One advantage of the APBMs and our estimation procedure is the capability of maintaining the physical interpretability of estimated states. Furthermore, we propose a constraint filtering approach to control the neural network contributions to the overall model. We also exploit assumed density filtering techniques and cubature integration rules to present a flexible estimation strategy that can easily deal with nonlinear models and high-dimensional latent spaces. Finally, we demonstrate the efficacy of our methodology by leveraging a target tracking scenario with nonlinear and incomplete measurement and acceleration models, respectively.

ABSTRACT. State estimation in non-linear models is performed by tracking the posterior distribution recursively. A plethora of algorithms have been proposed for this task. Among them, the Gaussian particle filter uses weighted sets of particles to construct Gaussian approximations to the posterior. In this paper, we propose to use invertible particle flow methods, derived under the Gaussian boundary conditions for a flow equation, to generate a proposal distribution close to the posterior. The resultant particle flow Gaussian particle filter (PFGPF) algorithm retains the asymptotic properties of Gaussian particle filters, with the potential for improved state estimation performance in high-dimensional spaces. We compare the performance of PFGPF with the particle flow filters and particle flow particle filters in two challenging numerical simulation examples.

ABSTRACT. In this paper the issue of filtering and smoothing in continuous discrete time is studied when the state variable evolves in some submanifold of Euclidean space, which may not have the usual Lebesgue measure. Formal expressions for prediction and smoothing problems are reviewed, which agree with the classical results except that the formal adjoint of the generator is different in general. These results are used to generalise the projection approach to filtering and smoothing to the case when the state variable evolves in some submanifold that lacks a Lebesgue measure. The approach is used to develop projection filters and smoothers based on the von Mises–Fisher distribution, which are shown to be outperform Gaussian estimators both in terms of estimation accuracy and computational speed in simulation experiments involving the tracking of a gravity vector.

ABSTRACT. Simultaneous localization and mapping (SLAM) using multipath at mmWave frequencies can provide accurate localization in the presence of static landmarks. When radio signals reflected from moving vehicle scatterers (VSs) are observed, the standard SLAM filters exhibit a degraded performance. To address this problem, we propose a probability hypothesis density (PHD)-based SLAM filter, where separate maps are maintained for VS and other types of landmarks. Through a combination of developed (i) local PHD-SLAM and (ii) global map fusion, we demonstrate that the proposed filter can handle the problem.

ABSTRACT. We consider the problem of jointly estimating the states of multiple global navigation satellite system (GNSS) receivers modeled with shared biases. In particular, we explore how to best assign these receivers to disjoint sets, so as to retain computational feasibility in the resulting filters. We propose a genetic algorithm that dynamically assigns agents to clusters subject to constraints on the maximum number of states in the clusters. Several numerical examples illustrate the flexibility of the approach, and the choice of genetic operations in the clustering algorithm is motivated by their effect on the algorithm's expected convergence rate. Numerical experiments with a GNSS-inspired problem demonstrates that the proposed clustering can yield a substantial improvements in the mean-square error compared to a random cluster assignment.

ABSTRACT. This paper discusses a recently developed efficient solution to the nonlinear filtering problem that arises in terrain-aided navigation of land and air vehicles. The specific implementation of a Rao-Blackwellised point-mass filter has been field-tested and its performance has been validated on multiple data sets and maps of variable accuracy. The filter implementation for the air vehicles is supplemented with a radar altimeter model that predicts the location of the most probable radar beam reflection point to be used in measurement consistency test.

ABSTRACT. Face biometrics has become an integral part of the various security and law enforcement applications, including border control scenarios. However, the face recognition systems are vulnerable to the morphing attacks, and thus, it is essential to develop a reliable and robust face Morphing Attack Detection (MAD) techniques. This paper presents a novel approach based on the residual gradients computed from the face image's colour scale-space representation in the reference-based or differential set-up. Thus, the proposed method will take two facial images (one from the passport and another from the trusted device) to compute the residual gradients, which is then classified using Spectral Regression Kernel Discriminant Analysis (SRKDA) to reliable detect the face morphing attacks. Extensive experiments are carried out on two different datasets to benchmark the performance of the proposed method, especially to different morph generation methods, morphing data mediums (digital, print-scan and print-scan compression) and ageing variations. Experimental results demonstrate the improved performance of the proposed method over the state-of-the-art reference-based face MAD in all evaluation protocols.

ABSTRACT. In the traditional methods of maneuvering target tracking, it is necessary to adjust the state transition model in time to match the maneuvering target, which will cause the problems of model decision delay and competition. Besides, the commonly adopted first-order Markov assumption can lead to the loss of information when motion modes are relevant to time. In order to solve these problems, a data-driven algorithm based on LightGBM is proposed in this paper. Maneuvering target tracking is modeled as a non-probabilistic method of direct mapping from sensor measurement to target state, track samples of different motion modes are used for training, and fast online tracking is realized. Comparing it with interacting multiple model (IMM) algorithm in a variety of different scenarios, simulation results show that the proposed method has advantages in accuracy and speed. Finally, the robustness of the algorithm is verified under the compound noise of Cauchy and Gaussian distributions.

ABSTRACT. The problem of fixed-interval smoothing for jump Markov systems is considered in this paper. An optimal backward-time recursive equation for computing the joint posterior of the state vector and model index is established first. A suboptimal algorithm is then developed to approximate the newly obtained Bayesian smoother under nonlinear state-space models with additive Gaussian noise. The proposed method utilizes the well-known assumed density filtering with Gaussian assumption and the expression for the quotient of two Gaussian densities to compute the smoothing posteriors. It eliminates the need for finding the inverse of the state dynamics and can handle singular process noise covariance, compared with several existing multiple model smoothers. Promising results are observed in simulations using a maneuvering target tracking task.

ABSTRACT. This paper studies the localization accuracy of a target in urban environment with line-of-sight conditions to the sensors based on acoustic bearing measurements. Instead of a commonly used wave-based approach, we focus on an information-theoretic analysis and derive the Cramér-Rao bound based on bearing measurements of direct wave and first-order reflections to predict the achievable target localization accuracy. Through this approach, we observe an improvement in accuracy by accounting for additional acoustic bearing measurements of reflected waves. We validate the match between predicted and actual performance in Monte Carlo simulations and field experiments in an urban environment. For this purpose, we propose an estimator that extends the classical bearings-only localization with bearing measurements of first-order reflections. Finally, we use a Cramér-Rao bound based optimization criterion, that accounts for direct and reflected waves, to optimize sensor placement in urban environment.

ABSTRACT. The process of combining data and estimates is inherent in estimation problems. This paper focuses on the linear fusion under the assumption that only some elements of the cross-correlation matrix of the estimation errors are known. Configurations of the knowledge are discussed individually for up to five estimates. For an arbitrary number of estimates, a general construction of upper bounds of the joint mean square error matrix is proposed. Last, the relation with the Split Covariance Intersection fusion is discussed.

ABSTRACT. Abstract—In this paper, a robust labeled multi-Bernoulli (RLMB) filter for the multi-target tracking (MTT) scenarios with inaccurate and time-varying process and measurement noise covariances is proposed. The process noise covariance and measurement noise covariance are modeled as inverse Wishart (IW) distributions, respectively. The state together with the predicted error and measurement noise covariances are inferred based on the VB inference. Moreover, a closed-form implementation of the proposed RLMB filter is given for linear Gaussian system and the predictive likelihood function is calculated by minimizing the Kullback-Leibler (KL) divergence by the VB lower bound. Simulation results illustrate that the proposed RLMB filter outperforms the existing LMB filter in the tracking performance.

ABSTRACT. Spatiotemporal imaging has applications in e.g. cardiac diagnostics, surgical guidance, and radiotherapy monitoring, In this paper, we explain the temporal motion by identifying the underlying dynamics, only based on the sequential images. Our dynamical model maps the inputs of observed high-dimensional sequential images to a low-dimensional latent space wherein a linear relationship between a hidden state process and the lower-dimensional representation of the inputs holds. For this, we use a conditional variational auto-encoder (CVAE) to nonlinearly map the higher dimensional image to a lower-dimensional space, wherein we model the dynamics with a linear Gaussian state-space model (LG-SSM). The model, a modified version of the Kalman variational auto-encoder, is end-to-end trainable, and the weights, both in the CVAE and LG-SSM, are simultaneously updated by maximizing the evidence lower bound of the marginal likelihood. In contrast to the original model, we explain the motion with a spatial transformation from one image to another. This results in sharper reconstructions and the possibility of transferring auxiliary information, such as segmentation, through the image sequence. Our experiments, on cardiac ultrasound time series, show that the dynamic model outperforms traditional image registration in execution time, to a similar performance. Further, our model offers the possibility to impute and extrapolate for missing samples.

ABSTRACT. With an increasingly ageing population in the world, fall detection and classification for elderly people becomes an imperative problem that needs to be addressed for assisted living. Currently, most of the fall detection algorithms are based on wearable and non-wearable sensors, such as based on accelerometer and video camera respectively. In this work, different from previous vision-based methods where the whole images are used, to mitigate the privacy protection problem and detect different types of fall events, we utilize only the skeleton data to achieve the classification of different fall events by using a deep neural network (DNN). The cost of manually labelling and due to varieties of annotators, for a recorded dataset, there always exist errors which will deteriorate the performance. To address this issue, we introduce the confident learning to remove wrong labelled samples and propose a new cascaded learning method to solve the noisy labelled data problem. To confirm the efficacy of the proposed method, we compare different algorithms on the UP-Fall dataset to show that the proposed method performs better than the state-of-the-art.

ABSTRACT. Numerous sleep disorders are characterised by movement during sleep, these include rapid-eye movement sleep behaviour disorder (RBD) and periodic limb movement disorder. The process of diagnosing movement related sleep disorders requires laborious and time-consuming visual analysis of sleep recordings. This process involves sleep clinicians visually inspecting electromyogram (EMG) signals to identify abnormal movements. The distribution of characteristics that represent movement can be diverse and varied, ranging from brief moments of tensing to violent outbursts. This study proposes a framework for automated limb-movement detection by fusing data from two EMG sensors (from the left and right limb) through a Dirichlet process mixture model. Several features are extracted from 10 second mini-epochs, where each mini-epoch has been classified as 'leg-movement' or 'no leg-movement' based on annotations of movement from sleep clinicians. The distributions of the features from each category can be estimated accurately using Gaussian mixture models with the Dirichlet process as a prior. The available dataset includes 36 participants that have all been diagnosed with RBD. The performance of this framework was evaluated by a 10-fold cross validation scheme (participant independent). The study was compared to a random forest model and outperformed it with a mean accuracy, sensitivity, and specificity of 94%, 48%, and 95%, respectively. These results demonstrate the ability of this framework to automate the detection of limb movement for the potential application of assisting clinical diagnosis and decision-making.

ABSTRACT. In this work, we propose a novel multisource deep learning architecture that employs the evidential Transferable Belief Model (TBM) for combining classifiers for Covid diagnosis. Our architecture was used in the difficult task of distinguishing mild cases of Covid versus severe ones that require urgent medical attention. The available datasets comprised radiographies and clinical data of the patients that we classified separately with a CNN and a decision tree respectively. In our approch, TBM was used systematically used to fuse both the the results of individual layers in the CNN and to combine the outputs of the CNN with the decision tree. The results outperform the individual classifiers and classical fusion methods.

ABSTRACT. This paper explores the current machine learning based methods used to identify Attention Deficit Hyperactivity Disorder (ADHD) in humans. With ADHD being one of the most prevalent mental health disorders worldwide, machine learning could be one of the effective solutions to objective diagnosis support to clinicians. We explore the use of machine learning with different sensing techniques such as functional Magnetic Resonance Imagery (fMRI) and Electroencephalography (EEG). Moreover, we also explore other approaches to detect ADHD such as computer based tasks, medical questionnaires and medical notes. With mental health awareness on the rise, it is necessary to survey the existing literature on ADHD for a machine learning based reliable Artificial Intelligence (AI). Which can aid clinicians in order to speed up the ADHD diagnosis process.

ABSTRACT. In networked estimation architectures, event-based sensing and communication can contribute to a more efficient resource allocation, in general, and improved utilization of communication resources, in particular. In order to tap the full potential of event-based scheduling, the design of transmission triggers and estimators need to be closely coupled while two directions are auspicious: First, the remote estimator can exploit the absence of transmissions and translate it into implicit information about the sensor data. Second, an intelligent trigger mechanism at the sensor that predicts future sensor readings can decrease transmission rates while rendering the implicit information more valuable. Such an intelligent trigger has been developed in a recent paper based on a Finite Impulse Response filter. This scheme requires the sensor to transmit an additional estimate alongside the measurement. In the present paper, the communication demand is further reduced by only transmitting the estimate. The remote estimator exploits correlations to incorporate the received information. In doing, the estimation quality is also improved, which is confirmed by simulations.

ABSTRACT. Situation-aware technologies enabled by multiobject tracking (MOT) methods will create new services and applications in fields such as autonomous navigation and applied ocean sciences. Belief propagation (BP) is a state-of-the-art method for Bayesian MOT but fully relies on a statistical model and preprocessed sensor measurements. In this paper, we establish a hybrid method for model-based and data-driven MOT. The proposed neural enhanced belief propagation (NEBP) approach complements BP by information learned from raw sensor data with the goal to improve data association and to reject false alarm measurements. We evaluate the performance of our NEBP approach for MOT on the nuScenes autonomous driving dataset and demonstrate that it can outperform state-of-the-art reference methods. In particular, the proposed NEBP method is currently in second place in the nuScenes LiDAR tracking challenge.

ABSTRACT. Reliable tracking algorithms are essential for automated driving. However, the existing consistency measures are not sufficient to meet the increasing safety demands in the automotive sector. Therefore, this work presents a novel method for self-assessment of single-object tracking in clutter based on Kalman filtering and subjective logic. A key feature of the approach is that it additionally provides a measure of the collected statistical evidence in its online reliability scores. In this way, various aspects of reliability, such as the correctness of the assumed measurement noise, detection probability, and clutter rate, can be monitored in addition to the overall assessment based on the available evidence. Here, we present a mathematical derivation of the reference distribution used in our self-assessment module for our studied problem. Moreover, we introduce a formula that describes how a threshold should be chosen for the degree of conflict, the subjective logic comparison measure used for the reliability decision making. Our approach is evaluated in a challenging simulation scenario designed to model adverse weather conditions. The simulations show that our method can significantly improve the reliability checking of single-object tracking in clutter in several aspects.

ABSTRACT. This paper deals with the problem of tracking multiple objects, in which each object is known to belong to a unique class. We follow the tracking by detection paradigm and assume that the object detector provides scores in addition to each detection. The problem is tackled as simultaneous classification and tracking using random finite sets. Inspired by the multi-Bernoulli mixture (MBM) filter, we propose a solution to the problem by modifying the target birth process. To simplify the implementation and to mitigate the computational costs, we develop tractable solutions with linear complexity. The algorithms are subsequently used for visual tracking of surgical instruments. As a side-product, we derive the prediction step of the Bernoulli filter using the probability generating functionals (PGFLs).

ABSTRACT. Advances in multi-object tracking (MOT) provide possibilities for unmanned monitoring across intelligent transportation.The tracking-by-detection paradigm aims at associating detection boxes whose scores are higher than a threshold with active trajectories. However, when there is no valid detection associated with certain trajectories due to occlusions or false negatives (FN) directly leads to fragmented trajectories and ID switches. An effective trajectory recovery method is proposed for multi-vehicle tracking in traffic surveillance in this paper. The proposed method provides a set of pseudo-feature embeddings to fill the gaps in the fragmented trajectories. These embeddings with motion and appearance features can be re-associated with new detections through a cascaded matching strategy that improves the quality of data association. Experimental results prove that our method can effectively reduce the number of ID switches and fragmentations. It is worth noting that the proposed method without a deep model achieves high tracking speed with considerable accuracy compared to state-of-the-art methods.

ABSTRACT. In nonlinear recursive Bayesian state estimation, it is in general intractable to maintain the exact posterior density for the state. In order to obtain an approximate solution, it is common to (re-)approximate the posterior density with a Gaussian density. In this paper, we investigate the use of deterministic samples generated by herding in Gaussian filters. Herding is a (relatively) computationally cheap procedure to sequentially generate deterministic samples from probability density functions. We demonstrate the practicability of herding in sample-based nonlinear Kalman filters and Gaussian particle filters in two challenging tracking scenarios.

ABSTRACT. We present a quasi-Monte Carlo acceptance-rejection sampling method for arbitrary multivariate continuous probability density functions. The method employs either a uniform or a Gaussian proposal distribution. The proposal samples are provided by optimal deterministic sampling based on the generalized Fibonacci lattice. By using low-discrepancy samples from generalized Fibonacci lattices, we achieve a more locally homogeneous sample distribution than random sampling methods for arbitrary continuous densities such as the Metropolis-Hastings algorithm or slice sampling, or acceptance-rejection based on state-of-the-art quasi-random sampling methods like the Sobol or Halton sequence.

ABSTRACT. The desensitized Kalman filter can robustly estimate the state of a system with uncertain parameters without knowledge about uncertainty type. In this paper, the desensitized Kalman filter for nonlinear systems is derived using Taylor series expansion and a stochastic approach to reduce estimation error sensitivity to uncertain parameters. Adaptively normalized weights tune the trade-off between the minimum uncertainty sensitivity and minimum mean square error. Among the main benefits of the algorithm are intuitive tuning concerning uncertainty and a form resembling the classical Riccati equation. The comparison to other robust state-of-the-art algorithms is discussed based on a numerical example.

ABSTRACT. Nonlinear estimation can be performed in many ways, with the particle filter being one of the most common. It is well known that ignoring the latest measurement value in the choice of importance density can result in poor particle filter performance for certain classes of challenging problems. In this paper a novel particle filter with importance density based on the linear minimum mean square error (LMMSE) estimator is presented. Performance is evaluated using Monte Carlo simulations of a highly nonlinear growth model. The proposed algorithm is compared to existing particle filter formulations using metrics for accuracy, consistency, and particle diversity.

ABSTRACT. The availability of a vast quantity of information from news channels and social media, make it often difficult to find and follow specific events. This applies to both causal readers and to intelligence and emergency response analysts. In particular, the latter need to find and process relevant information within sense-making, situation and impact assessment processes. The automatic retrieval and tracking of news has been addressed by a good number of works in the information retrieval literature. However, there is a strong potential for introducing automatic systems employing information fusion methods and techniques to assist decision makers. In the field of deep learning, several techniques for text encoding have been proposed, which have allowed significant progress also in the field of news retrieval and ranking. The purpose of this paper is to explore the usage and combination of different pre-trained sentence embeddings, including multimodal ones, obtained from different parts of text that compose a news item. This in order to understand which type of technique is best for encoding the different information available on online news.

ABSTRACT. Hybrid threat events are rare and cannot be modelled solely based on data. Instead they require a focus on discovery of emergent knowledge through information sharing across agencies and systems. That requires a shared conceptualisation of the problem and entities involved. It also means that uncertain and possibly conflicting information describing multiple entities and their relationships needs to be reasoned about.

In this paper we discuss the relationship between uncertain conceptual graphs and belief functions. We put forward a fusion process which allows for taking advantage of evidential reasoning capabilities in a multi-entity context.

We show how information from conceptual graphs can be fed into or represented as an evidential networks and how the inference results obtained from valuation networks can be used to generate a probability distribution on conceptual graphs. This is demonstrated on a multi-entity threat assessment situation where a hybrid threat is formed by several possibly cooperating vessels.

ABSTRACT. Fusion of high-level symbolic reasoning with lower level signal-based reasoning has attracted significant attention. We propose an architecture that integrates the high-level symbolic domain knowledge using a hierarchical planner with a lower level reinforcement learner that uses hybrid data (structured and unstructured). We introduce a novel neuro-symbolic system, Hybrid Deep ReRPeL that achieves the best of both worlds---the generalization ability of the planner with the effective learning ability of deep RL. Our results in two domains demonstrate the superiority of our approach in terms of sample efficiency as well as generalization to increased set of objects.

ABSTRACT. n order to protect their common heritage of culture, personal freedom, and the rule of law in an increasingly fragile world, democracies must be able “to fight at machine speed” if necessary. For this reason, digitization in defense cannot not be confined to logistics, maintenance, intelligence, surveillance, and reconnaissance, but must equally enable responsible weapons engagement. We therefore try to contribute to a timeless problem: How to decide ‘well’ according to what is recognized as ‘true’? Turned into systems engineering, this leads to three tasks of responsible controllability: (1) Desing AI-based automation in a way that human beings are not only mentally but also emotionally able to master each situation. (2) Identify technical design principles to facilitate the responsible use of AI-based automation in defence. (3) Guarantee that human decision makers have full superiority of information, decision-making, and options of action. The Ethical AI Demonstrator (E-AID) proposed is paving the way to aswers by letting soldiers experience the use of AI in a military scenario with all associated aspects of psychological stress as realistically as possible.

ABSTRACT. We introduce a new approach to characterize and detect threatening behaviors in surveillance systems, without relying on history or expertise. This approach consists in simulating the worst-case attack plans, fusing their semantic descriptions and using the produced patterns to raise alerts in operational conditions. We demonstrate our set of tools on a simple scenario involving geolocated sensors looking for moving vehicles targeting a protected objective. We find that the system is able to recover well-grounded graph patterns defining detection rules which make sense in the operational context. We believe that our approach achieves a relevant compromise between data-based and expertise-based systems, and allows for a good balance between efficiency and understandability.

ABSTRACT. Coordinate systems and coordinate system conversions for bistatic range, bistatic range-rate, and azimuthal angle measured by a surface-wave radar are presented. Expressions for the Cramer-Rao Lower Bound (CRLB) for error analysis are provided and are demonstrated in a measurement conversion scenario.

ABSTRACT. In this paper we develop a method for vehicle positioning based on global navigation satellite system (GNSS) and camera information. Both GNSS and camera measurements have noise characteristics that vary in time. As a result, the measurements can abruptly change from reliable to unreliable from one time step to another. To adapt to the changing noise levels and hence improve positioning performance, we combine GNSS information with measurements from a forward looking camera, a steering-wheel angle sensor, wheel-speed sensors, and optionally an inertial sensor. We pose the estimation problem in an interacting multiple-model (IMM) setting and use Bayes recursion to choose the best combination of the estimators. In a simulation study, we comapre vehicle models with varying complexity, and on a real road segment we show that the proposed method can accurately adjust to changing noise conditions.

ABSTRACT. Fusing thermal and visible images is a recurring challenge in computer vision, especially when the images of the two modalities are not well registered. This registration problem is traditionally solved by matching descriptors and depends on the richness and discriminating power of the representation. Ensuring that detected features are dense and uniformly distributed is not necessarily guaranteed. More recently, machine learning methods addressed the issue of visible to visible matching, but few address the multi-modality setting. In this paper, we propose to address the special case of thermal-visible image registration with small baseline parallax correction. Our deep homography model is evaluated on an open thermal and visible dataset with two training settings, unsupervised and supervised. Results demonstrate the feasibility of the approach, and performances comparison to state-of-the-art models is evaluated.

ABSTRACT. This paper considers meta-cognition for cognitive radars in an adversarial setting. A cognitive radar optimally adapts its waveform (response) in response to maneuvers (probes) of a possibly adversarial moving target. A meta-cognitive radar is aware of the adversarial nature of the target and seeks to mitigate the adversarial target. How should the meta-cognitive radar choose its responses to sufficiently confuse the adversary trying to estimate the radar's utility function? This paper abstracts the radar's meta-cognition problem in terms of the spectra (eigenvalues) of the state and observation noise covariance matrices, and embeds the algebraic Riccati equation into an economics-based utility maximization setup. This adversarial target is an inverse reinforcement learner. By observing a noisy sequence of radar's responses (waveforms), the adversarial target uses a statistical hypothesis test to detect if the radar is a utility maximizer. In turn, the meta-cognitive radar deliberately chooses sub-optimal responses that increasing its Type-I error probability of the adversary's detector. We call this counter-adversarial step taken by the meta-cognitive radar as inverse inverse reinforcement learning (I-IRL). We illustrate the meta-cognition results of this paper via simple numerical examples. Our approach for meta-cognition in this paper is based on revealed preference theory from micro-economics and inspired by results in differential privacy and adversarial obfuscation in machine learning.

ABSTRACT. Targetless Lidar-camera registration is a repeating task in many computer vision and robotics applications and requires computing the extrinsic pose of a point cloud with respect to a camera or vice-versa. Existing methods based on learning or optimization lack either generalization capabilities or accuracy. Here, we propose a combination of pre-training and optimization using a neural network-based mutual information (MI) estimation technique (MINE). This construction allows back-propagating the gradient to the calibration parameters and enables stochastic gradient descent. To ensure orthogonality constraints with respect to the rotation matrix we incorporate Lie-group techniques. Furthermore, instead of optimizing on entire images, we operate on local patches that are extracted from the temporally synchronized projected Lidar points and camera frames. Our experiments show that this technique not only improves over existing techniques in terms of accuracy, but also shows considerable generalization capabilities towards new Lidar-camera configurations.

ABSTRACT. This paper presents a new effective measure of un- certainty (MoU) of basic belief assignments. This new continuous measure is effective in the sense that it satisfies a small number of very natural and essential desiderata. Our new simple mathematical definition of MoU captures well the interwoven link of randomness and imprecision inherent to basic belief assignments. Its numerical value is easy to calculate. This new effective MoU characterizes efficiently any source of evidence used in the belief functions framework. Because this MoU coincides with Shannon entropy for any Bayesian basic belief assignment, it can be also interpreted as an effective generalization of Shannon entropy. We also provide several examples to show how this new MoU works.

ABSTRACT. This paper considers the integrated localization for the mixed near-field (NF) and far-field (FF) sources using the uniform linear array (ULA). With the help of the polynomial rooting methods and the propagator, an efficient algorithm is proposed to provide an integrated estimation of the direction of arrival (DOA) and the ranges of the sources. It takes low computational burden without the requirements that separating the DOA and range information or pre-classification of the sources. We first construct two special fourth-order cumulant matrices using the received array data, then extract the prior-electrical parameters related to the array elements by its steering matrix, and finally carry out parameter matching and classification. Besides, the proposed algorithm eliminates the need for tedious eigenvalue decomposition and spectral search steps, and has almost no aperture loss. Eventually, several simulation results show that the proposed algorithm has lower computational complexity under an acceptable accuracy, compared to the state-of-the-art methods.

ABSTRACT. Small-Scale Unmanned Autonomous Systems (sUASs) have become an integral part of defense and homeland security domains for numerous situation assessment and decision support tasks. In selecting sUASs for these applications, evaluation methods that provide decisive comparisons of different sUASs platforms is critical for not only selecting appropriate sUASs, but also for making sure the required minimum capabilities are met for a given mission. Toward this, a new quantitative Operator Situational Awareness (OSA) assessment method is presented for evaluating sUASs for their ability to provide adequate levels of OSA in subterranean (SubT) environments. The work presented in this paper improves on our prior work by applying formulas to the SEEV model for accurate quantification and experiments with new operationally relevant scenarios designed to validate the formulas. The OSA assessment method presented in this paper is an improvement of our previous work based on Attention Allocation Model (AAM) and Man-Machine Integration Design and Analysis (MIDAS)-based SA model. In addition, we apply this improved method to other platforms in order to use this method on various platforms rather than the specific sUAV. In this paper, we present a method by accounting for spatial perception differences across platforms by introducing a new component that we refer to as Virtual Proportion (VP), which is obtained from AAP of other other SEs by comparing the correct rate of Situation Awareness Global Assessment Technique (SAGAT). We compare 2 military-grade sUAS platforms and discussed how to expand the test method and apply it.

ABSTRACT. Navigation using a Global Navigation Satellite System (GNSS) is common for autonomous vehicles (ground or air). Unfortunately, GNSS-based navigation solutions are often susceptible to jamming, interference, and a limited number of satellites. A proposed technique to aid in navigation when a GNSS-based system fails is magnetic navigation - navigation using the Earth's magnetic anomaly field. This solution comes with its own set of problems including the need for quality magnetic maps in every area in which magnetic navigation will be used. Many of the currently available magnetic maps are generated from a combination of dated magnetic surveys which use less-sensitive equipment to track the magnetic anomaly field. The resulting maps are riddled with spatially correlated errors the structure of which is largely unknown. The correlations are further confounded while navigating because they depend on how fast a vehicle moves through the map in addition to the original correlated error structure. Traditionally, this spatial correlation has been handled by introducing a First Order Gauss-Markov (FOGM) noise model into the estimation routine, with the FOGM parameters set somewhat arbitrarily. In this paper, we investigate the possibility of using correlation agnostic fusion techniques (i.e., Covariance Intersection and Probablisticly Conservative Fusion) for magnetic navigation. These techniques have the advantage of not requiring any parameter tuning; the same method and tuning parameters are used regardless of the spatial correlation. We demonstrate that utilizing probabilisticly conservative fusion leads to navigation results that are better than many tuned approaches and reasonably close to the best possible tuning parameters of a FOGM.

ABSTRACT. Passive sensors pose challenges in the localisation and tracking of targets due to the inherent ambiguity in the source range—specifically, accurate 2-D localisation using only a single sensor is not guaranteed. On the other hand, the underwater domain exhibits certain characteristics that induce some structure to the problem that help reduce the range ambiguity and lead to a favourable 2-D localisation and tracking performance. This work incorporates target motion analysis (TMA) into tracking algorithms by specifying target birth process parameters and demonstrates that the output exhibits contraction in the range uncertainty. In addition, we propose a likelihood model that incorporates cepstrum processing detections that exploit multi-path reflections from the sea surface and the sea bottom to reduce the range ambiguity. We use these models in the generalised labelled multi-Bernoulli (GLMB) multi-object model and the associated tracking filter with sequential Monte Carlo techniques. This model explicitly incorporates track labels in Bayesian recursions yielding coherent trajectory estimates consisting of the 2-D location estimates when used with passive detections.