FUSION 2022: 25TH INTERNATIONAL CONFERENCE ON INFORMATION FUSION (FUSION 2022)
PROGRAM FOR THURSDAY, JULY 7TH
Days:
previous day
all days

View: session overviewtalk overview

10:30-12:10 Session 13A: SS12: Localization for Autonomous Systems - Part I
Location: Operan
10:30
Decentralized Collaborative Localization with Map Update using Schmidt-Kalman Filter
PRESENTER: Maxime Escourrou

ABSTRACT. This paper presents a new decentralized approach for collaborative localization and map update relying on landmarks measurements performed by the robots themselves. The method uses a modified version of the Kalman filter, namely Schmidt Kalman filter that approaches the performance of the optimal centralized Kalman filter without the need to update each robot pose. To deal with data incest and limited communication, the computation of  cross-covariance errors between robots must be well managed. Each robot individually updates its own map, the map fusion is performed by using the unweighted Kullback-Leibler Average to keep estimation consistency.The performance of the approach is evaluated in a simulation environment where robots are equipped with odometry and a lidar for exteroceptive perception. The results show that collaboration improves the localization of the robots and the estimation of the map while maintaining consistency.

10:50
A Geometric Approach to Passive Localisation

ABSTRACT. In this paper, we present a geometric framework for the passive localisation of static emitters. The objective is to localise the position of the emitters in a given area by centralised coordination of mobile passive sensors. This framework uses only the geometry of the problem to minimise the maximal bounds of the emitters' locations without using a belief or probability distribution. This geometric approach provides effective boundaries on the emitters' position. It can also be useful in evaluating different decision-making strategies for coordinating mobile passive sensors and complementing statistical methods during the initialisation process. The effectiveness of the geometric approach is shown by designing and evaluating a greedy decision-making strategy, where a sensor selects its future position by minimising the maximum uncertainty on its next measurement using one of the global objective functions. Finally, we analyse and discuss the emergent behaviour and robustness of the proposed algorithms.

11:10
Sensor Fusion and Optimal Platform Trajectory Planning for Ground Target Localization with Terrain Uncertainty and Measurement Biases

ABSTRACT. A ground target can be localized using a biased airborne angle-only sensor. However, the measurement bias causes a delay in error convergence. Adding measurements from an unbiased range sensor improves localization by attaining faster convergence. However, planning an optimal trajectory for the range sensor platform is a challenging problem. To ensure observability in 3-D localization, in most of the recent works, the height of the ground target from the sea level is assumed to be known perfectly. However, in most practical applications, target height is obtained from a Digital Terrain Elevation Database (DTED), having multiple levels of resolution. As a result, in addition to the bias uncertainty, the terrain uncertainty is required to be handled. Cramer Rao Lower Bound (CRLB) is derived for the localization problem, considering the terrain and measurement bias uncertainties. We also propose a CRLB based optimization algorithm for optimal range sensor platform trajectory planning. In this work, we proposed two localization approaches. The first approach involves bias compensation using a prior whereas the target and bias states are estimated jointly in the second approach. In both approaches, range sensor fusion is proposed to improve localization accuracy. The effectiveness of our algorithms is verified using Monte Carlo simulations.

11:30
Experimental Validation of Single Base Station 5G mmWave Positioning: Initial Findings
PRESENTER: Yu Ge

ABSTRACT. 5G cellular networks can utilize millimeter wave signals, and support large bandwidths and large antenna arrays, which provide more geometric-based signals and higher delay and angel resolutions. These merits bring new opportunities in positioning the user with limited infrastructure through the use of combined angle and delay information. However, there are many practical challenges to overcome, in order to have a functioning single base station 5G mmWave positioning system. In this paper, we describe a deployed single base station mmWave positioning system, and provide an example of the measurement data. Furthermore, we perform measurement validation on a limited measurement data set by performing base station localization. Additional evaluations performed on simulation model data provide guidelines on the required size of the data set and receiver antenna configuration, which will be implemented in upcoming measurements.

11:50
Fault tolerant cooperative localization using diagnosis based on Jensen Shannon divergence
PRESENTER: Zaynab El Mawas

ABSTRACT. Multi-robot systems have taken an important place in various applications, where at all times, the integrity of the location of robots must be ensured. This can be achieved by integrating a step of detecting and excluding sensor faults. In this article, a multi-sensor multi-vehicle Cooperative Positioning System (CPS) method tolerant to sensor faults is presented. The estimator used in this work is the informational form of the Kalman filter (KF), namely the Informational Filter (IF). To detect and isolate faults, the residuals generated are based on the divergence of Jensen Shannon (DJS ) between the probability distributions predicted and corrected by the IF. These residuals, as they are a sum of two divergences of Kullback-Leibler include two tests: one compares the means and the other compares the covariance matrices. For the optimization of the threshold of the residuals, the operating characteristic of the receiver (frequently referred to as the ROC curve) is used. The approach is tested on real data acquired by three Turtlebot3 equipped with wheel encoders (for odometry), a gyroscope (for the yaw angle), a Marvelmind localization system (for the position), and an Optitrack system (for ground truth).

10:30-12:10 Session 13B: SS7: Multiple Sensor Data Processing for Tracking, Classification and Intentionality Prediction - Part I
Location: Musikalen
10:30
A Learning Distributed Gaussian Process Approach for Target Tracking over Sensor Networks

ABSTRACT. Tracking manoeuvring targets often relies on complex models with non-stationary parameters. Gaussian process (GP) based model-free methods can achieve accurate performance in a data-driven manner but face scalability challenges. Aiming to address such challenges, this paper proposes a distributed GP-based tracking approach able to learn the kernel hyperparameters in an online manner, to improve the tracking performance and scalability. It caters to the inherent distributed feature of sensor networks and does not need measurements to be transmitted among sensors for target states predictions. Theoretical upper confidence bounds about the tracking error are derived within the regret bound setting. Through this theoretical analysis, the tracking error per time step is upper bounded as a function of predictive variances from local sensors. The theoretical results are supported by simulation based ones over a case study for tracking ever wireless sensor networks. With evaluation on challenging target trajectories, a comparison on state-of-the-art centralised and distributed GP approaches, numerical results demonstrate that the proposed approach achieves competitively high and robust tracking performance.

10:50
A Variational Bayes Association-based Multi-object Tracker under the Non-homogeneous Poisson Measurement Process
PRESENTER: Runze Gan

ABSTRACT. The non-homogeneous Poisson process (NHPP) has been widely used to model extended object measurements where one target can generate zero or several measurements; it also provides an elegant solution to the computationally demanding data association problem in multiple target tracking. This paper presents an association-based NHPP system, coupled with which we propose a variational Bayes association-based NHPP (VB-AbNHPP) tracker that can estimate online the target kinematics and the association variables in parallel. In particular, the VB-AbNHPP tracker can be easily extended to include online static parameter learning (e.g. measurement rates) based on a general coordinate ascent variational filtering framework developed here. The results show that the proposed VB-AbNHPP tracker is superior to other competing methods in terms of implementation efficiency and in tracking accuracy.

11:10
An Improved Two-Stage Based Multi-frame Track-Before-Detect Algorithm in Radar systems
PRESENTER: Xiujuan Lu

ABSTRACT. The detection and tracking of dim targets in complex environments show great importance in radar systems. The track-before-detect (TBD) technology has been widely researched in the scenario where the target signal-to-noise ratio (SNR) is low, however, there has been no suitable solution to the trade-off between tracking accuracy and computational complexity during the process of multi-frame joint detection in radar systems. In this paper, we propose an efficient two-stage based multi-frame detection and tracking algorithm in radar systems. The proposed algorithm consists of a low threshold pre-processing stage, and a TBD processor, which searches possible target tracks from multiple scans and declares the final estimated tracks. The proposed algorithm provides an accurate evolution of target states over time in polar coordinates to avoid the performance loss and model mismatch due to the nonlinear conversion in mixed coordinates. In addition, we further propose a greedy-based recursive algorithm to implement fast track formation from the over-threshold multi-frame measurement points. Simulation results show that the proposed method achieves a better detection and tracking performance with a low computational complexity.

11:30
Addressing data association by message passing over graph neural networks

ABSTRACT. In cooperative localization systems, the fusion of information from multiple sensing platforms is acknowledged to improve localization accuracy of sensed targets. However, the data association required to perform the inference is non-trivial to be solved. In this context, we propose a graph formulation of the data association problem among unlabelled information produced at different sensors in which we run a Message Passing Neural Network (MPNN). The proposed MPNN algorithm suits for centralized sensing architecture where all sensors are connected to a single processing unit. We validate the theoretic aspects with numerical simulations in a vehicular scenario with cooperative lidar sensing. We show the robustness of the model against several environmental complexities such as high number of cooperative vehicles and different noise intensities.

11:50
Self-Assessment and Robust Anomaly Detection with Bayesian Deep Learning
PRESENTER: Nidhal Bouaynaya

ABSTRACT. Deep Learning (DL) models have achieved or even surpassed human-level accuracy in several areas, including computer vision and pattern recognition. The state-of-art performance of DL models has raised the interest in using them in real-world applications, such as disease diagnosis and clinical decision support systems. However, the challenge remains the lack of trustworthiness and reliability of these DL models. The detection of incorrect decisions or flagging suspicious input samples is essential for the reliability of machine learning models. Uncertainty estimation in the output decision is a key component in establishing the trustworthiness and reliability of these models. In this work, we use Bayesian techniques to estimate the uncertainty in the model's output and use this uncertainty to detect distributional shifts linked to both input perturbations and labels shifts. We use the learned uncertainty information (i.e., the variance of the predictive distribution) in two different ways to detect anomalous input samples: 1) a static threshold based on average uncertainty of a model evaluated on the clean test data, and 2) a statistical threshold based on the significant increase in the average uncertainty of the model evaluated on corrupted (anomalous) samples. Our extensive experiments demonstrate that both approaches can detect anomalous samples. We observe that the proposed thresholding techniques can distinguish misclassified examples in the presence of noise, adversarial attacks, anomalies or distributional shifts. For example, when considering corrupted versions of MNIST and CIFAR-10 datasets, the rate of detecting misclassified samples is almost twice as compared to Monte-Carlo-based approaches.

10:30-12:10 Session 13C: Bayesian Inference and Decision Theory
Location: Verdefoajén
10:30
Recursive Joint Cramér-Rao Lower Bound for Nonlinear Parametric Systems with Colored Noise
PRESENTER: Xianqing Li

ABSTRACT. The performance evaluation for joint state and parameter estimation (JSPE) is of great significance. Joint Cramér-Rao lower bound (JCRLB) has been widely studied for JSPE of nonlinear parametric systems with white noise. However, in practice, the noise is often colored due to high measurement frequency and bandlimited signal channels. In this paper, a recursive JCRLB is developed for JSPE of nonlinear parametric systems with colored noise, characterized by auto-regressive (AR) models. First, we propose a unified recursive JCRLB for JSPE of general nonlinear parametric systems with higher-order autocorrelated process noises and autocorrelated measurement noise simultaneously. Then its relationship with the posterior Cramér-Rao lower bound (PCRLB) for filtering of nonlinear systems with colored noise and the hybrid Cramér-Rao lower bound (HCRLB) for JSPE of regular parametric systems with white noise are provided. Illustrative examples in radar target tracking verify the effectiveness of the proposed JCRLB for the performance evaluation for JSPE of nonlinear parametric systems with colored noise.

10:50
On the Observability of Gaussian Models using Discrete Density Approximations
PRESENTER: Ariane Hanebeck

ABSTRACT. This paper proposes a novel method for testing observability in Gaussian models using discrete density approximations (deterministic samples) of (multivariate) Gaussians. Our notion of observability is defined by the existence of the maximum a posteriori estimator. In the first step of the proposed algorithm, the discrete density approximations are used to generate a single representative design observation vector to test for observability. In the second step, a number of carefully chosen design observation vectors are used to obtain information on the properties of the estimator. By using measures like the variance and the so-called local variance, we do not only obtain a binary answer to the question of observability but also provide a quantitative measure.

11:10
Uncertainty Aware Quickest Change Detection: An Experimental Study
PRESENTER: Lance Kaplan

ABSTRACT. In this work, we study the problem of Quickest Change Detection which aims to detect when a stream of observations transitions from being drawn from a pre-change distribution to a post-change distribution as quickly as possible. Traditionally, either information is completely known about the distributions, or no information is known and their parameters are estimated using frequentist approaches, e.g., Generalized Likelihood Ratio test. Recently, the Uncertain Likelihood Ratio (ULR) test was proposed for the QCD problem which relaxes both of these assumptions to form a Bayesian test that allows for no knowledge, partial knowledge, and full knowledge of the parameters of the distributions. In this work, we extend the ULR test to improve the order of operations required to compute the test statistic using a windowing method to form the Windowed Uncertain Likelihood Ratio (W-ULR) algorithm. We then applied it to multivariate Gaussian observations and empirically evaluated the average detection delay and missed detections for various false alarm rates under various operating conditions. The results show that the W-ULR outperforms the (windowed) GLR test, which is consistent with the initial findings.

11:30
SOLBP: Second-Order Loopy Belief Propagation for Inference in Uncertain Bayesian Networks
PRESENTER: Federico Cerutti

ABSTRACT. In second-order Bayesian networks, network conditional probabilities are known within distributions. The delta-method has been applied to extend exact first-order inference methods to propagate both means and variances through sum-product networks derived from Bayesian networks, thereby characterizing epistemic uncertainty. Second-order belief propagation, instead, has been demonstrated for polytrees but not for general directed acyclic graph structures. In this work, we extend Loopy Belief Propagation to the setting of second-order Bayesian networks, giving rise to Second-Order Loopy Belief Propagation (SOLBP). SOLBP generates inferences consistent with those generated by sum-product networks while being more computationally efficient and scalable.

11:50
Fast optimize-and-sample method for differentiable Galerkin approximations of multi-layered Gaussian process priors
PRESENTER: Simo Särkkä

ABSTRACT. Multi-layered Gaussian process (field) priors are non-Gaussian priors, which offer a capability to handle Bayesian inference on both smooth and discontinuous functions. Previously, performing Bayesian inference using these priors required the construction of a Markov chain Monte Carlo sampler. To converge to the stationary distribution, this sampling technique is computationally inefficient and hence the utility of the approach has only been demonstrated for small canonical test problems. Furthermore, in numerous Bayesian inference applications, such as Bayesian inverse problems, the uncertainty quantification of the hyper-prior layers is of less interest, since the main concern is to quantify the randomness of the process/field of interest. In this article, we propose an alternative approach, where we optimize the hyper-prior layers, while inference is performed only for the lowest layer. Specifically, we use the Galerkin approximation with automatic differentiation to accelerate optimization. We validate the proposed approach against several existing non-stationary Gaussian process methods and demonstrate that it can significantly decrease the execution time while maintaining comparable accuracy. We also apply the method to an X-ray tomography inverse problem. Due to its improved performance and robustness, this new approach opens up the possibility for applying the multi-layer Gaussian field priors to more complex problems.

10:30-12:10 Session 13D: Methods for Classification and Decision Making
Location: Spegelsalen
10:30
Paraphrasing Techniques for Maritime QA system
PRESENTER: Fatemeh Shiri

ABSTRACT. There has been an increasing interest in incorporat- ing Artificial Intelligence (AI) into Defence and military systems to complement and augment human intelligence and capabilities. However, much work still needs to be done toward achieving an effective human-machine partnership. This work is aimed at enhancing human-machine communications by developing a capability for automatically translating human natural language into a machine-understandable language (e.g., SQL queries). Techniques toward achieving this goal typically involve building a semantic parser trained on a very large amount of high-quality manually-annotated data. However, in many real-world Defence scenarios, it is not feasible to obtain such a large amount of training data. To the best of our knowledge, there are few works trying to explore the possibility of training a semantic parser with limited manually-paraphrased data, in other words, zero-shot. In this paper, we investigate how to exploit paraphrasing methods for the automated generation of large-scale training datasets (in the form of paraphrased utterances and their corresponding logical forms in SQL format) and present our experimental results using real-world data in the maritime domain

10:50
Unsupervised Clustering of Marine Vessel Trajectories in Historical AIS Database
PRESENTER: R. Praveen Jain

ABSTRACT. This paper presents results on unsupervised clustering of marine vessel trajectories contained in a database of historical AIS messages. The historical AIS database is a collection of individual AIS messages from multiple vessels over a specific time period. Inherent in these messages are vessel trajectories that define a set of marine traffic patterns. Clustering these trajectories in an unsupervised manner help obtaining clustered datasets pertaining to unique maritime traffic lanes, which serves as an input to the situational awareness system of marine vessels. The clustering algorithm employed is non parametric and proceeds in three stages i) Trajectory extraction that pre-processes the AIS database to extract vessel trajectories, ii) Trajectory shrinking that uses Manifold-blurring Mean-Shift algorithm in combination with Principal Component Analysis applied locally to the dataset, in order to increase similarity between trajectories belonging to the same cluster, and iii) Trajectory Clustering that clusters the shrunk trajectories using the Longest Common Sub-sequence trajectory similarity measure and Agglomerative Hierarchical Clustering algorithm. The clustering algorithm exploits the inherent structure between the position and course over ground measurements contained within the AIS messages and its efficacy are demonstrated using historical AIS data collected over Trondheimsfjørd. Further, the utility of the clustered dataset is demonstrated through a case-study of long term trajectory prediction of a target vessel.

11:10
On the Theoretical Convergence and Error Sensitivity Analysis of Yayambo for Fusion of Probabilistic Classifier Outputs

ABSTRACT. This paper considers the problem of classifier fusion for situations where classifier outputs are probabilistic. We study the convergence and analyse the error sensitivity of a performance-agnostic fusion of probabilistic classifier outputs from \cite{masakuna2020performance} (called Yayambo). This method is iterative and was proposed for combining probabilistic outputs of black-box classifiers trained for the same task to make a single consensus class prediction. Yayambo considers the diversity between the outputs of the various classifiers, iteratively updating predictions based on their correspondence with other predictions until the predictions converge to a consensus decision or the number of iterations has exceeded a prefixed threshold. For this paper, we address two things. First, as an iterative fusion process, convergence is an ideal way to find consensus \cite{lorenz2010conditions}. Cases for convergence in the original method was shown experimentally. We establish its conditional theoretical convergence. There may be various reasons to study the convergence of an iterative algorithm, including insurance of obtention of a consensus decision. It should be noted that consensus class does not necessarily mean true class, i.e. Yayambo can converge to a wrong class. Here, the main motivation is the following: investigating on the theoretical convergence of Yayambo might lead into obtaining a closed form expression for the consensus class from the initial probabilities without performing the iteration explicitly. We apply the squeeze theorem \cite{sohrab2003basic} to prove convergence of Yayambo and find that Yayambo converges under some conditions. Finally, individual classifiers' outputs always contain estimation errors, which can impact the fusion decision. Following \cite{kittler1998combining, kittler1998combining1}, we analyse the error sensitivity of Yayambo and find that the error factor can have a dramatic impact on the consensus class. Analysing error sensitivity of a classifier fusion model describes how errors in classifier outputs can (negatively) impact the outcome of the classifier fusion model. This error sensitivity analysis can indicate, for some situations where one needs to select a classifier fusion method for example, which fusion methods would be more appropriate than others.

11:30
Classification and Fusion of Two Disparate Data Streams and Nuclear Dissolutions Application
PRESENTER: Nageswara Rao

ABSTRACT. We consider two streams of data or measurements with disparate qualities and time resolutions that need to be classified. The first stream consists of higher quality data at a coarser time resolution, and the other consists of lower quality data at a finer time resolution. We present a fuser-switch method that fuses the set of classifiers of each stream separately and switches between them. We show that this method provides classification decisions at a finer time resolution with superior detection and false alarm probabilities compared to individual classifiers, under the statistical independence and time resolution ratio conditions. When classifiers are trained using machine learning methods, we show that this superior performance is guaranteed with a confidence probability specified by the classifiers' generalization equations. We use these results to provide analytical foundations for previous practical results that achieved significant performance improvements in classifying Pu/Np target dissolution events at a radiochemical processing facility.

11:50
LiDAR-Landmark Modeling for Belief-Space Planning using Aerial Forest Data
PRESENTER: Jonas Nordlöf

ABSTRACT. A belief-space planning problem for GNSS-denied areas is studied, where knowledge about the landmark density is used as prior, instead of explicit landmark positions. To get accurate predictions of the future information gained from observations, the probability of detecting landmarks needs to be taken into account in addition to the probability of the existence of landmarks. Furthermore, these probabilities need to be calculated from prior data without knowledge of explicit landmarks. It is shown in this paper how the landmark detection probabilities can be generated for a ground-to-ground LiDAR sensor and integrated in the path-planning problem. Moreover, it is also shown how prior information can be generated for a forest scenario. Lastly, the approach is evaluated in a simulated environment using a real landmark detector applied to a simulated point cloud. Compared to previous approaches, an informative path planner, integrating the proposed approximation, is able to reduce the platform pose uncertainty. This is achieved using only prior aerial data of the environment.

13:40-15:20 Session 14A: SS7: Multiple Sensor Data Processing for Tracking, Classification and Intentionality Prediction - Part II
Location: Musikalen
13:40
Detection of outliers in classification by using quantified uncertainty in neural networks

ABSTRACT. Neural Networks (NNs) can solve very hard classification and estimation tasks, but are less well suited to solve complex sensor fusion challenges, such as end-to-end control of autonomous vehicles. NN can still be a powerful tool for particular sub-problems in sensor fusion. But this would require a reliable and quantifiable measure of the stochastic uncertainty in the predictions that can be compared to classical sensor measurements. Current NN's output some figure of merit, which is, however, only a relative model fit and not a stochastic uncertainty.

We propose to embed the NN's in a proper stochastic system identification framework. In the training phase, the stochastic uncertainty of the parameters in the (last layers of the) NN is quantified. We show that this can be done recursively with very little extra computations. In the classification phase, Monte-Carlo (MC) samples are used to generate a set of classifier outputs. From this set, a distribution of the classifier output is obtained, which represents a proper description of the stochastic uncertainty of the predictions. We also show how to use the calculated uncertainty for outlier detection by including an artificial outlier class. In this way, the NN fits a sensor fusion framework much better.

We evaluate the approach on images of handwritten digits. The proposed method is shown to be on par with MC dropout, while having lower computational complexity, and the outlier detection almost completely eliminates false classifications.

14:00
Deep Learning for Audio Visual Emotion Recognition

ABSTRACT. Human emotions can be presented in a variety of ways and can involve multiple modalities, hence developing an automated system to identify human emotions is a challenging task. Deep neural architectures have been used and shown excellent performance to predict emotional states in recent years. In this work, we propose an emotion recognition system based on speech and visual modalities, where visual cues can convey the speech/non-speech frames and emotional content/state of the speaker. Different from previous research works, we propose to use a Convolutional Neural Network (CNN) to extract Wavegram and Wavegram-Logmel features from speech similar to log mel spectrogram, but learned using a neural network to utilise the information from both time-domain waveforms and log mel spectrograms. The system is then trained in an end-to-end fashion where – by also taking advantage of the correlations of each of the streams – we manage to outperform the traditional and state-of-the-art deep neural-based emotion recognition approaches based on auditory and visual handcrafted features for the prediction of spontaneous and natural emotions on the SAVEE database.

14:20
Efficient Factorisation-based Gaussian Process Approaches for Online Tracking

ABSTRACT. Target tracking often relies on complex models with non-stationary parameters. Gaussian process (GP) is a model-free method that can achieve accurate performance. However, the inverse of the covariance matrix poses scalability challenges. Since the covariance matrix is typically dense, direct inversion and determinant evaluation methods suffer from cubic complexity to data size. This bottleneck limits the GP for long-term tracking or high-speed tracking. We present an efficient factorisation-based GP approach without any additional hyperparameters. The proposed approach reduces the computational complexity of the Cholesky decomposition by hierarchically factorising the covariance matrix into off-diagonal low-rank parts. Meanwhile, the resulting low-rank approximated Cholesky factor can also reduce the computation complexity of the inverse and the determinant operations. Numerical results based on offline and online tracking problems demonstrate the effectiveness of the proposed approach.

14:40
UAV-enabled Edge Computing for Optimal Task Distribution in Target Tracking

ABSTRACT. Unmanned aerial vehicles (UAVs) are useful devices due to their great manoeuvrability for long-range outdoor target tracking. However, these tracking tasks can lead to sub-optimal performance due to high computation requirements and power constraints. To cope with these challenges, we design a UAV-based target tracking algorithm where computationally intensive tasks are offloaded to Edge Computing (EC) servers. We perform joint optimization by considering the trade-off between transmission energy consumption and execution time to determine optimal edge nodes for task processing and reliable tracking. The simulation results demonstrate the superiority of the proposed UAV-based target tracking on the known trajectory over several existing techniques.

15:00
Notes on the Product Multi-Sensor Generalized Labeled Multi-Bernoulli Filter and its Implementation
PRESENTER: Martin Herrmann

ABSTRACT. The product multi-sensor generalized labeled multi-Bernoulli filter constitutes a multi-object filter for centralized and distributed multi-sensor systems with centralized estimator. It implements the Bayes parallel combination rule for generalized labeled multi-Bernoulli densities, simplifying the NP-hard multi-dimensional k-best assignment problem of the multi-sensor multi-object update to a polynomial-time k-shortest path problem. This way, the filter allows for efficient, parallelizable, and distributed calculation of the multi-sensor multi-object update, while showing excellent performance. However, the derivation of the filter formulas relies on a well-established approximation of the fundamental multi-sensor Gaussian identity, which was inadvertently not labeled as such in our original article [1]. Thus, on the one hand, we clarify this mistake, discuss its consequences, and present a mathematically clean derivation of the filter yet to establish the claim of Bayes-optimality. On the other hand, we discuss implementation details and present extensive evaluations, that complete the previous publication of the filter.

13:40-15:20 Session 14B: SS12: Localization for Autonomous Systems - Part II
Location: Operan
13:40
Trilateration Using Motion Models

ABSTRACT. In this paper, we present a framework for doing localization from distance measurements, given an estimate of the local motion. We show how we can register the local motion of a receiver, to a global coordinate system, using trilateration of given distance measurements from the receivers to senders in known positions. We describe how many different motion models can be formulated within the same type of registration framework, by only changing the transformation group. The registration is based on a test and hypothesis framework, such as RANSAC, and we present novel and fast minimal solvers that can be used to bootstrap such methods. The system is tested on both synthetic and real data, with promising results.

14:00
Using Probabilistic Geometrical Map Information For Train Localization
PRESENTER: Wendi Löffler

ABSTRACT. Determination of train positions within a railway network must be fail-safe and of high accuracy. This is an essential task to solve to achieve a secure and efficient railway operation. In this paper, we present a method to estimate position and velocity of a train in the track net using given position estimates from an arbitrary information source, and improving the estimate by using geometrical track information. We focus on modelling and exploiting of the geometrical track information including possible uncertainties and examine the impact of uncertainties on the state estimate.We store the track information as a set of supporting points with Gaussian uncertainties and interpolate linearly. The track information is fed into a Kalman filter in form of soft constraints that is modified to account for state-dependent observation noise. A simulated test run shows that the average position and velocity error along track decreases significantly when modelling the uncertainty of the constraints, compared to using a Kalman filter with hard constraints. We evaluate the presented filter for different supporting point and measurement uncertainties and show that the performance within a typical parameter setting for train positioning is improved compared to the unconstrained Kalman filter and the Kalman filter with hard constraints.

14:20
Error-State Unscented Kalman-Filter for UAV Indoor Navigation

ABSTRACT. In this paper, we present an algorithm for indoor quadcopter navigation. We implemented a strapdown navigation algorithm combined with an error-state unscented Kalman-Filter being capable of fusing IMU, barometer and UWB measurements. Optical flow and distance to ground measurements are additionally fused to further improve the state estimation quality. Compared to alternate approaches, the suggested algorithm has better trajectory following abilities and does not rely on the actual quadcopter's dynamics, so it can be applied to a variety of flying platforms. We implemented and evaluated the algorithm on the Crazyflie 2.1 mini-quadcopter.

14:40
Feature Based Multi-Hypothesis Map Representation for Localization in Non-Static Environments
PRESENTER: Kristin Nielsen

ABSTRACT. Long-term autonomy of robots requires localization in an inevitably changing environment, where the robots' knowledge about the surroundings are more or less uncertain. Inspired by methods in target tracking, this paper proposes a multi-hypothesis feature based map representation to provide robust localization under these conditions. It is derived how this representation can be used to obtain consistent position estimates while at the same time providing up-to-date map information to be shared by cooperative robots or for visual presentation. Simulations are performed that conceptually highlights the benefit of the developed solution in an environment where uniquely identifiable landmarks are moved between discrete positions. This relates to a real world scenario where a robot moves in a corridor with office doors opened or closed at different times.

15:00
Visual-Inertial Odometry aided by Speed and Steering Angle Measurements
PRESENTER: Andreas Serov

ABSTRACT. Visual-inertial navigations systems (VINS) can be an essential building block for autonomous systems, that are equipped with a camera and an inertial measurement unit (IMU), especially in indoor or GNSS-denied environments. One prevalent visual-inertial odometry (VIO) algorithm is the multi-state constraint Kalman Filter (MSCKF). OpenVINS is an open source project that performs visual-inertial state estimation using the core functionality of the MSCKF. However, it offers several extensions. It is designed for general purpose, but in certain cases it is beneficial to use domain knowledge in order to increase accuracy and robustness in state estimation. In this paper, we propose using vehicle speed, steering angle, and wheel speeds measurements in conjunction with OpenVINS by performing additional filter updates in the context of autonomous driving. The updates are conducted using a classical Kalman filtering approach, where speed and steering measurements are processed at their respective sensor frequency. Additionally, all measurements between camera frames, which have the slowest measurement frequency, are gathered and a 3 degrees of freedom (DOF) planar motion update is performed in a preintegrated fashion. In contrast to handheld devices and drones commonly used as visual-inertial platforms, the movement of automotive vehicles is constrained in a nonholonomic way which is reflected in the updates. These extensions are evaluated on real-world datasets of typical urban driving. The results show that state estimation with the additional updates is smoother and leads to lower translation errors, where a preintegrated vehicle update using a single-track model offers the best overall performance. The code is provided in a fork of the original project: https://github.com/aserbremen/open_vins

13:40-15:20 Session 14C: SS4: Data Fusion for Industry 4.0
Location: Spegelsalen
13:40
Efficient Task Allocation in Smart Warehouses with Multi-Delivery Stations and Heterogeneous Robots
PRESENTER: Juha Röning

ABSTRACT. Robotic Mobile Fulfillment Systems (RMFS) are crucial for increasing the efficiency of the logistics sector. One of the essential processes in such systems is efficiently allocating tasks to robots to reduce costs and speed up order fulfillment. In this context, we found a gap in exploring the possibility of using heterogeneous robot fleets in RMFS plus other real constraints, such as dynamic orders and multi-delivery stations. This work presents a space decomposition-driven heuristic based on Voronoi Tessellation concepts to efficiently allocate tasks to robots in an environment with several constraints. Experiments deal with a large simulated smart warehouse, with dynamic orders and thousands of tasks and robots. Results showed that our solution reduced route costs and time to fulfill orders up to 45%  and 24%, respectively, compared to the state-of-the-art algorithm.

14:00
Sensor Fusion for Detection and Localization of Carbon Dioxide Releases for Industry 4.0
PRESENTER: Gianluca Tabella

ABSTRACT. This work tackles the distributed detection & localization of carbon dioxide (CO2) release from storage tanks caused by the opening of pressure relief devices via inexpensive sensor devices in an industrial context. A realistic model of the dispersion is put forward in this paper. Both full-precision and rate-limited setups for sensors are considered, and fusion rules capitalizing the dispersion model are derived. Simulations analyze the performance trends with realistic system parameters (e.g. wind direction).

14:20
Fusion of Sentiment and Asset Price Predictions for Portfolio Optimization

ABSTRACT. The fusion of public sentiment data in the form of text with stock price prediction is a topic of increasing interest within the financial community. However, the research literature seldom explores the application of investor sentiment in the Portfolio Selection problem. This paper aims to unpack and develop an enhanced understanding of the sentiment aware portfolio selection problem. To this end, the study uses a Semantic Attention Model to predict sentiment towards an asset. We select the optimal portfolio through a sentiment-aware Long Short Term Memory (LSTM) recurrent neural network for price prediction and a mean-variance strategy. Our sentiment portfolio strategies achieved on average a significant increase in revenue above the non-sentiment aware models. However, the results show that our strategy does not outperform traditional portfolio allocation strategies from a stability perspective. We argue that an improved fusion of sentiment prediction with a combination of price prediction and portfolio optimization leads to an enhanced portfolio selection strategy.

14:40
Intrinsic and Extrinsic Calibration Method for a Trinocular Multimodal Camera Setup
PRESENTER: Carsten Plasberg

ABSTRACT. Machine Vision and its common usage in automation processes as well as autonomous systems make cameras essential parts of many modern robotic applications. With cameras always comes the challenge of calibration, as e.g. lenses and sensor distort the image. The use of several cameras can be motivated by either redundancy, a wider field of view or the usage of different camera technologies to improve vision capabilities. Still those cameras need an extrinsic calibration so adjacent points in images of different cameras can be linked. This paper describes a method to calibrate a multimodal camera setup with simple features that can be detected by all types of cameras used in the setup. The proposed method leads to an easily applicable routine that, beside being used in laboratory environments, also can be used in field and addresses both intrinsic and extrinsic calibration for a trinocular setup with a thermal, an event-based and a combined color/depth camera, all roughly facing in the same direction. Evaluation is done with a small combined mount for the cameras that can be used as a ground truth for extrinsic calibration while the proposed features can be used to determine quality of the intrinsic calibration with common methods.

15:00
Evaluation and comparison of eight popular Lidar and Visual SLAM algorithms

ABSTRACT. In this paper, we evaluate eight popular and open-source 3D Lidar and visual SLAM (Simultaneous Localization and Mapping) algorithms, namely LOAM, Lego LOAM, LIO SAM, HDL Graph, ORB SLAM3, Basalt VIO, and SVO2. We have devised experiments both indoor and outdoor to investigate the effect of the following items: i) effect of mounting positions of the sensors, ii) effect of terrain type and vibration, iii) effect of motion (variation in linear and angular speed). We compare their performance in terms of relative and absolute pose error. We also provide a comparison of their required computational resources. We thoroughly analyze and discuss the results and identify the best performing system for the environment cases presented in our multi-camera and multi-Lidar indoor and outdoor datasets. We hope our findings help one to choose a sensor and the corresponding SLAM algorithm combination suiting their needs, based on their target environment.

13:40-15:20 Session 14D: Stochastic Processes
Location: Verdefoajén
13:40
Gaussian Process Regression with Grid Spectral Mixture Kernel: Distributed Learning for Multidimensional Data

ABSTRACT. Kernel design for Gaussian processes (GPs) along with the associated hyper-parameter optimization is a challenging problem. In this paper, we propose a novel grid spectral mixture (GSM) kernel design for GPs that can automatically fit multidimensional data with affordable model complexity and superior modeling capability. To alleviate the curse of dimensionality, we leverage a multicore computing environment to optimize the kernel hyper-parameters in a distributed manner. We further propose a doubly distributed learning algorithm based on the alternating direction method of multipliers (ADMM) which enables multiple agents to learn the kernel hyper-parameters collaboratively. The doubly distributed learning algorithm is shown to be effective in reducing the overall computational complexity while preserving data privacy during the learning process. Experiments on various one-dimensional and multidimensional data sets demonstrate that the proposed kernel design yields superior training and prediction performance compared to its competitors.

14:00
Continuous Model Evaluation and Adaptation to Distribution Shifts: A Probabilistic Self-Supervised Approach
PRESENTER: Gregor Pavlin

ABSTRACT. This paper introduces a Bayesian approach to estimating distribution shifts over the modelled variables and continuous model adaptations to mitigate the impact of such shifts. The method exploits probabilistic inference over sets of correlated variables in causal models describing data generating processes. By extending the models with latent auxiliary variables, probabilistic inference over sets of correlated variables enables estimation of the distribution shifts impacting different parts of the models. Moreover, the introduction of latent auxiliary variables makes inference more robust against distribution shifts and supports automated, self-supervised adaptation of the modelling parameters during the operation, often significantly reducing the adverse impact of the distribution shifts. The effectiveness of the method has been validated in systematic experiments using synthetic data.

14:20
Intention-Aware Motion Modeling Using GP Priors With Conditional Kernels
PRESENTER: Linfeng Xu

ABSTRACT. For the challenging task of modeling actual complex motions, we propose a new class of Gaussian process (GP)models that are data-driven and also take into account prior knowledge of the motion intention. As a theoretical basis, we show that the GP regression is mathematically equivalent to regularized least-squares estimation for random functions with known prior means. Compared with the popular GP models in machine learning literature, the proposed GP motion model priors with conditional kernels have at least two advantages: 1) they are nonstationary and more applicable to represent complex motions by integrating the basic kinematic principles; 2) conditional kernels are further devised by incorporating the motion intent so that the resultant GP models are more versatile and would expectedly entail more accurate trajectory prediction. The properties used for efficient computation are studied. Finally, illustrative examples are provided to show the superiority of the proposed motion models and to verify the theoretical results given in the paper.

14:40
Gaussian Approximation Filter Based on Divergence Minimization for Nonlinear Dynamic Systems
PRESENTER: Liping Guo

ABSTRACT. In the Bayesian filtering paradigm, approximation of posterior distribution is an important research topic for nonlinear dynamic systems. In this paper, we aim at obtaining its Gaussian approximation via KL divergence minimization. We formulate the problem as a nonlinear programming problem with linear constraints and resort to the feasible direction method for the solution. Since the gradient of the objective function involves intractable integrals, we adopt a cubature rule to calculate the gradient, which is suitable for real-time filtering for its simplicity, efficiency, and accuracy. Based on the Gaussian approximation, a nonlinear filter is derived, and it is demonstrated to be effective by simulations.

15:00
Deep Learning For Online Time Series Trend Prediction
PRESENTER: Morgan Blem

ABSTRACT. Recent studies show that Deep Neural Networks can be highly effective for trend prediction applications. These studies, however, typically focus on offline applications. In this study we explore deep neural networks (DNN) for trend prediction in real time online applications. We reformulate the trend prediction problem for online applications, and present an efficient online trend segmentation algorithm to update the current trend, which is used to predict the how long until the next trend change and the slope of the next trend. Four DNN algorithms, i.e. LSTMs, CNNs, BiLSTMs and TCNs are implemented and evaluated across four different datasets using walk-forward validation. The recurrent DNN models, specifically the BiLSTM, outperforms the other algorithms. The findings suggest that DNNs can be effectively used for online trend prediction in real-time applications.