Program

Program Overview

 

 

All talks are 10 minutes + 5 minutes Q&A. The names of the talks are listed in the table below the session name. You can look up the authors and abstracts of the papers below. All times are in Central European Summer Time (CEST).
September 23 First Conference Day
Room A (Zoom Link) (Maps Link) Room B (Zoom Link) (Maps Link)
11:00 – 12:30 Lunch Vogelbräu
13:00 – 14:00 Plenary Talk: Davide Scaramuzza
14:00 – 14:30 Coffee Break
14:30 – 15:30

Robotic Vision and Systems (1)

  • Latency Analysis of ROS2 Multi-Node Systems
  • On-Demand Virtual Highways for Dense UAS Operations
  • Skeleton-based Fall Events Classification with Data Fusion
  • Automatic Identification of the Leader in a Swarm using an Optimized Clustering and Probabilistic Approach

Directional Estimation (1)

  • Bingham-Gaussian Distribution on S^3 x R^n for Unscented Attitude Estimation
  • Mean Mixtures in Directional Statistics
  • Series expansion and Pfaffian systems of the von Mises distribution on the torus
  • An Inequality for Bayesian Bregman Risks With Applications in Directional Estimation
15:30 – 16:00 Coffee Break
16:00 – 17:00

Robotic Vision and Systems (2)

  • IMU-based pose-estimation for spherical robots with limited resources
  • Towards Robust VSLAM in Dynamic Environments: A Light Field Approach
  • Robust multi-stage hybrid vision/force control of industrial robots
  • Detection of Conductive Lane Markers using mmWave FMCW Automotive Radar

Directional Estimation (2)

  • Directional DD-classifiers under non-rotational symmetry
  • Conditional Densities and Likelihoods for Densities Based on Trigonometric Polynomials for Hypertori
  • Time-efficient Bayesian inference for a (skewed) von Mises distribution on the torus in a Deep Probabilistic Programming Language
  • Bayesian inference for skew-wrapped Cauchy mixture model using a modified Gibbs sampler
17:00 – 17:20 Break
17:20 – 18:35

Tracking

  • Shape Estimation and Tracking using Spherical Double Fourier Series for Three-Dimensional Range Sensors
  • Incorporating Range-Rate Measurements in EKF-based Elliptical Extended Object Tracking
  • Object Detection and Mapping with Bounding Box Constraints
  • Adiabatic Quantum Computing for Solving the Multi-Target Data Association Problem
  • Observability-Aware Trajectory Optimization: Theory, Viability & State of the Art

Machine Learning

  • A Convolutional Neural Network Combined with a Gaussian Process for Speed Prediction in Traffic Networks
  • Generalization Equations for Machine Learners Based on Physical and Abstract Laws
  • CS-Fnet: A Compressive Sampling Frequency Neural Network for Simultaneous Image Compression and Recognition
  • Evaluation of Image Classification Networks on Impulse Sound Classification Task
19:00 Welcome Reception: Hoepfner Burghof

 

September 24 Second Conference Day
Room A (Zoom Link) (Maps Link) Room B (Zoom Link) (Maps Link)
11:00 – 12:30 Lunch Badische Weinstuben
13:00 – 14:00 Plenary Talk: Darius Burschka
14:00 – 14:30 Coffee Break
14:30 – 15:45

Multisensor Systems and Fusion

  • Lower Bounds In Estimation Fusion With Partial Knowledge of Correlations
  • Learning and Exploiting Partial Knowledge in Distributed Estimation
  • Audio-Video Sensor Fusion for the Detection of Security Critical Events in Public Spaces
  • A Question of Time: Revisiting the Use of Recursive Filtering for Temporal Calibration of Multisensor Systems
  • Noninvasive Continuous Tracking of Partial Pressure of Oxygen in Arterial Blood: Adapting Microorganisms Bioprocess Soft Sensor Technology for Holistic Analysis of Human Respiratory System

Localization and Navigation

  • Optimal Sensor Placement for Shooter Localization within a Surveillance Area
  • LiDAR-Aided Relative and Absolute Localization for Automated UAV-based Inspection of Aircraft Fuselages
  • Robot Localization in a Pipe Network using a Particle Filter with Error Detection and Recovery in a Hybrid Metric-Topological Space
  • α-Rényi based framework for a Robust and Fault-Tolerant Localization
  • Navigation of a Self-Driving Vehicle Using One Fiducial Marker
15:45 – 16:15 Coffee Break
16:15 – 17:30

Nonlinear Estimation

  • Estimation of wheel odometry model parameters with improved Gauss-Newton method
  • A Novel Gamma Filter for Positive Parameter Estimation
  • UKF Parameter Tuning for Local Variation Smoothing.
  • Kalman Filtered Compressive Sensing Using Pseudo-Measurements
  • Gaussian Mixture Estimation from Weighted Samples

Perception

  • Point Clouds With Color: A Simple Open Library for Matching RGB and Depth Pixels from an Uncalibrated Stereo Pair
  • Mono-Vision based Moving Object Detection using Semantic-Guided RANSAC
  • Target-free Extrinsic Calibration of a 3D-Lidar and an IMU
  • SemCal: Semantic LiDAR-Camera Calibration using Neural Mutual Information Estimator
  • Segmentation of Stereo-Camera Depth Image into Planar Regions based on Evolving Principal Component Clustering
17:45 – 18:15 Award Ceremony
19:00 Conference Dinner: Badisch Brauhaus

 

Paper Details

Paper ID Paper Title Authors Abstract
1 Robust multi-stage hybrid vision/force control of industrial robots Bahar Ahmadi (Concordia University ); Wen-Fang Xie (Concordia University)*; Ehsan Mr Zakeri (Concordia University) This paper presents a novel scheme for robust hybrid vision/force control of industrial robots, subject to model uncertainties. It aims to improve the three phases of the control process: a) free-motion using the image-based visual servoing (IBVS) before the interaction with the workpiece; b) the moment that the end-effector touches the workpiece; and c) hybrid vision/force control during the interaction with the workpiece. The proposed method operates in three stages. First, the camera motion is decomposed into transitional and angular movements. Then, utilizing a switching method, the rotational and translational movements of the camera are controlled in the first two stages, respectively. In the last stage, hybrid vision/force control is activated. For each stage, super-twisting sliding mode controller (STSMC) is utilized. Employing STSMC results in robustness against uncertainties while addressing the chattering problem. A variable-gain sliding surface is also utilized to address the instability and convergence speed issues of the traditional switch IBVS. The experimental results demonstrate the effectiveness and superiority of the proposed multi-stage scheme compared to other traditional approaches.
2 Adiabatic Quantum Computing for Solving the Multi-Target Data Association Problem Felix Govaers (Fraunhofer FKIE)*; Veit Stooß (Fraunhofer FKIE); Martin Ulmke (Fraunhofer FKIE) Quantum computing promises significant improve- ments of computation capabilities in various fields such as machine learning and complex optimization problems. Recent technological advancements suggest that the adiabatic quantum computing ansatz may soon see practical applications. In this work, we adopt this computation paradigm to develop a quantum computation based solver of the well–known multi–target data association (MTDA) problem, an NP-hard nonlinear integer programming optimization task. The feasibility of the presented model is demonstrated by numerical simulation of the adiabatic evolution of a system of quantum bits towards the optimal solution encoded in the model Hamiltonian.
3 Navigation of a Self-Driving Vehicle Using One Fiducial Marker YIBO LIU (York University)*; Hunter Schofield (York University); JINJUN SHAN (York University) Navigation using only one marker, which contains four artificial features, is a challenging task since camera pose estimation using only four coplanar points suffers from the rotational ambiguity problem in a real-world application. This paper presents a framework of vision-based navigation for a self-driving vehicle equipped with multiple cameras and a wheel odometer. A multiple camera setup is presented for the camera cluster which has 360 degrees vision such that our framework solely requires one planar marker. A Kalman-Filter-based fusion method is introduced for the multiple-camera and wheel odometry. Furthermore, an algorithm is proposed to resolve the rotational ambiguity problem using the prediction of the Kalman Filter as additional information. Finally, the lateral and longitudinal controllers are provided. Experiments are conducted to illustrate the effectiveness of the theory.
4 Bingham-Gaussian Distribution on S^3 x R^n for Unscented Attitude Estimation Weixin Wang (The George Washington University)*; Taeyoung Lee (GW SEAS) In this paper, we propose the Bingham-Gaussian (BG) distribution defined on the Cartesian product of the three-sphere and the Euclidean space of an arbitrary dimension. BG is constructed by transforming the matrix Fisher-Gaussian distribution using the homomorphism between unit quaternions and the three dimensional rotation group. BG can be used to model the correlation between the attitude and the Euclidean space in a global fashion without singularities, and it properly deals with the cyclic nature of the space representing attitude, which is particularly advantageous when the uncertainty is large. An unscented attitude filter is designed based on BG which estimates the attitude and gyroscope bias concurrently. The proposed filter is compared against the conventional extended Kalman filter using numerical simulations.
5 Audio-Video Sensor Fusion for the Detection of Security Critical Events in Public Spaces Michael Hubner (AIT Austrian Institute of Technology)*; Christoph Wiesmeyr (AIT Austrian Institute of Technology GmbH); Klaus Dittrich (AIT Austrian Institute of Technology GmbH); Bernhard Kohn (AIT Austrian Institute of Technology GmbH); Heinrich Garn (AIT Austrian Institute of Technology GmbH); Martin Litzenberger (AIT Austrian Institute of Technology GmbH) Increasing security concerns in the public domain lead to a more widespread use of video and audio surveillance techniques. While both technologies are already advanced, they still produce high false alarm rates when deployed on their own under realistic conditions. We present a method for sensor fusion based on weighted maps and a rule engine. The system was tested in a public space with the combination of audio localization, audio classification and video crowd detection, using 79 simulated security relevant scenarios and 44 hours sample data recorded over a period of several weeks. It has been shown that the false positive rate was reduced by ~60% and the localization accuracy has been increased by 25% with the fusion approach compared to the detection performance of individual sensors alone.
9 Mean Mixtures in Directional Statistics Priyanka Nagar (University of Pretoria)*; Andriette Bekker (University of Pretoria); Mohammad Arashi (University of Mashhad) We construct a family of distributions on the unit hyper-sphere by conditioning a spherical normal mean mixture distribution with a general weight function. The resulting family of distributions accounts for symmetry, asymmetry, unimodality and bimodality depending on the choice of the weight function and parameter set. The problem of identifiability is partially addressed by re-parameterization. The circle and sphere are considered as special cases with various weight functions in a numerical illustration. A synopsis of the estimation procedure utilized in the numerical study is given. The fit of the proposed model to two real data sets as well as a simulated data set is evaluated and compared to well known circular models.
11 LiDAR-Aided Relative and Absolute Localization for Automated UAV-based Inspection of Aircraft Fuselages Roland Pugliese (RWTH Aachen University)*; Thomas Konrad (RWTH Aachen University); Dirk Abel (RWTH Aachen University) Robust localization is a crucial task for various
robotic applications, such as the inspection of aircraft surfaces
by autonomous Unmanned Aerial Vehicles (UAV). Currently,
aircraft inspection in mostly done in GNSS-denied hangar
environments but UAV indoor localization using external positioning
systems has several deficits regarding practicability,
safety and security. Since aircraft standing times are the crucial
cost factor, inspection should perspectively be shifted outdoors
to the runway in order to further increase efficiency, which
requires an onboard localization of the UAV relative to the
aircraft that ensures robustness by working independendly of
hangar features or external hardware. This paper proposes a
novel approach to exploit a-priori knowledge of a cylindrical
shaped inspection object, such as an aircraft fuselage, in order
to support and supplement the position state estimation of the
UAV. The relative localization approach comprises shape fitting
of a cylinder model into LiDAR measurements of the aircraft
using Random Sample Consensus (RANSAC), preceded by an
intensity-based prefiltering, resulting in a significant increase of
performance. This method can be applied to individual LiDAR
sweeps, enabling onboard computation in soft real-time with a
cycle time of 100 ms. Moreover, with the help of the onboard
LiDAR-derived relative positioning to the aircraft, the problem
of continuous line of sight to any transmitter of an indoor navigation
system (iGPS) is alleviated by integration into an existing
sensor fusion for absolute localization, while maintaining an
accuracy requirement in the submeter range. The feasibility of
the proposed approach is verified experimentally by evaluating
measurements obtained in a hangar with a Boeing 737-500.
13 Lower Bounds In Estimation Fusion With Partial Knowledge of Correlations Jiri Ajgl (University of West Bohemia)*; Ondrej Straka (University of West Bohemia) Mean square error matrices belong to key concepts in decentralised estimation. They assess the quality of estimates and are essential for the optimisation of the estimation fusion. In the case of a missing knowledge, sets of admissible matrices are replaced by their bounds and a robust fusion is applied. This paper prospects a specific partial knowledge of the sets of matrices. Upper bounds are constructed first. Then, the stress is laid on non-zero lower bounds, which do not exist in the fusion under completely unknown correlation. Limit cases are discussed on numerical examples and graphical illustrations are given.
15 A Novel Gamma Filter for Positive Parameter Estimation Corey L Marcus (The University of Texas at Austin)*; Renato Zanetti (The University of Texas at Austin) This work develops a new recursive Bayesian estimation for gamma distributed random variables. The time prediction and measurement update steps are developed, and both are shown to have analytic closed form solutions under certain conditions. Furthermore, the update is shown to be unbiased, this is theoretically guaranteed by the choice of estimator and demonstrated numerically through Monte Carlo methods with a simple example.
17 Series expansion and Pfaffian systems of the von Mises distribution on the torus Kazuya Suzuki (Kyoto University); Tomonari Sei (The University of Tokyo)* The normalizing constant of the von Mises distribution on a torus does not have a closed form, where the torus means the direct product of several circles. We derive a series expansion of the normalizing constant together with an error bound, which is useful for numerical computation. We also obtain a Pfaffian system in two-dimensional case, where the Pfaffian system is a set of differential equation necessary to perform the holonomic gradient method.
20 On-Demand Virtual Highways for Dense UAS Operations Thomas Henderson (University of Utah)*; David Sacharny (University of Utah); Vista Marston (University of Utah) The methods and protocols for coordinating airspace access can impact a number of metrics that are important to operators and system designers. For example, the time when an operator would like to fly a particular trajectory may be delayed if there are many intersecting trajectories by other aircraft. From a system-level perspective, it would be helpful to know how the number of simultaneous operations affects this measure of delay. Regarding the method of coordination, it is also important to consider what information must be shared in order to provide safe access – it may be undesirable to share detailed trajectory information. This paper describes a method for building on-demand airspace networks and applies Lane- Based Strategic Deconfliction (LBSD) to support dense airspace operations.
21 Noninvasive Continuous Tracking of Partial Pressure of Oxygen in Arterial Blood: Adapting Microorganisms Bioprocess Soft Sensor Technology for Holistic Analysis of Human Respiratory System Renaldas Urniezius (Kaunas University of Technology)*; Donatas Levisauskas (Kaunas University of Technology); Arnas Survyla (Kaunas University of Technology); Lina Jankauskaite (Lithuanian University of Health Sciences Hospital Kauno Klinikos); Goda Laucaityte (Lithuanian University of Health Sciences Hospital Kauno Klinikos); Dovile Lukminaite (Lithuanian University of Health Sciences Hospital Kauno Klinikos); Vygandas Vaitkus (Kaunas University of Technology) Early diagnosis of human life support fundamentally depends on dissolved oxygen in the arterial microvasculature. This study used respiratory and pressure data to noninvasively estimate the partial pressure of oxygen in the radial artery. Usually, the PaO2 sample analysis is an episodic action, and timely correction of hypoxemia requires convenient and accurate arterial monitoring of the arterial PaO2. Analysis of the blood samples confirmed that the proposed sensor fusion method has an average estimation error of 1.77 mm Hg of PaO2. The experiment duration was one hour and 46 minutes, and at the end of the trial, the bias fluctuation of estimation was about one mmHg of PaO2 when matched with the offline blood sample results. The sensor fusion approach involved the occasional 129.06 millibars of ambient overpressure applied to the human body. The preclinical trial also demonstrated the therapy effect that even relatively small overpressure, excluding head area, has a noticeable therapy effect when the healthy patient continuously inhales the 15-16% oxygen concentration air for one hour and 30 minutes. The investigation results are relevant for operating invasive and noninvasive lung ventilation and treating the current pandemic patients.
22 IMU-based pose-estimation for spherical robots with limited resources Jasper Zevering (University of Wuerzburg)*; Anton Bredenbeck (University of Wuerzburg) Spherical robots are a robot format that is not yet thoroughly studied for the application of mobile mapping. However, in contrast to other forms, they provide some unique advantages. For one, the spherical shell provides protection against harsh environments, e.g., guarding the sensors and actuators against dust and solid rock. This is particularly useful in space applications. Furthermore, the inherent rotation the robot uses for locomotion can be exploited to measure in all directions without having the sensor itself actuated. To use this rotation in combination with sensor data to create a consistent environment map, a reasonable estimation of the robot pose is required. In such cases the pose estimation can be done by inertial measurements from sensors such as accelerometers and gyroscopes, as interpolating instances for calculation-intensive slow algorithms as optical localization algorithms or as rough estimation. We propose a pose estimation procedure based on inertial measurements that exploits the known dynamics of a spherical robot. It emphasizes a low jitter to maintain constant world measurements during standstill and avoids exponentially growing error in position estimates. Evaluating the position and orientation estimates with given ground truth frames shows that we can reduce the jitter in orientation and handle slip and partly slide behavior as other commonly used filters such as the Madgwick filter.
23 Detection of Conductive Lane Markers using mmWave FMCW Automotive Radar Austin Greisman (Queen’s University)*; Keyvan Hashtrudi-Zaad (); Joshua Marshall (Queen’s University) Localization of vehicles in inclement weather conditions, including snow and heavy rain, is a significant issue plaguing autonomous vehicle systems. Our work takes a step towards tackling this problem by leveraging existing hardware commonly used in self-driving vehicles, namely low-cost millimetre wave (mmWave) radar systems to detect conductive paint on roads. This paper presents the results of preliminary experiments that indicated that, even with full snow converge of the conductive paint, the radar system is still able to successfully detect the material as a potential marker for lateral vehicle localization.
24 Point Clouds With Color: A Simple Open Library for Matching RGB and Depth Pixels from an Uncalibrated Stereo Pair Jordan R Nowak (LIRMM)*; Philippe Fraisse (LIRMM); Andrea Cherubini (U. Montpelier); Jean-Pierre Daures (Clinique Beau Soleil) Current day robots often rely – for visual perception – on the coupling of two cameras: one for color and one for depth. While for custom RGB-D cameras, the manufacturer takes care of aligning the two images, this is not done when two commercial cameras are coupled (e.g., on the Pepper robot) without having been calibrated beforehand. In this article, we present a simple open library for reconstructing the 3D position of RGB pixels without knowing the parameters of the two cameras. The library requires a simple preliminary calibration step based on pixel-to-pixel matching, and then automatically reconstructs 3D colored point clouds from a given set of pixels in the RGB image. The source code is available at the following link https://github.com/jordan-nowak/OpenHSML.
25 Bayesian inference for skew-wrapped Cauchy mixture model using a modified Gibbs sampler Najmeh Nakhaeirad (Department of Statistics, University of Pretoria)*; Andriette Bekker (University of Pretoria); Mohammad Arashi (Ferdowsi University of Mashhad) More and more datasets on the circle tend to exhibit non-trivial features such as
skewness and multimodality. Therefore, there is a growing demand for flexible models to capture these characteristics. For analyzing data with skew and multimodal patterns, we focus on the skew mixture models, specially skew-wrapped Cauchy mixture model as the underlying distribution.
We incorporate prior knowledge about the parameters into analysis to improve results.
A modified Gibbs sampling is used to obtain the estimates of the parameters. A simulation study is conducted to assess the performance of the proposed Bayesian approach for skew-wrapped Cauchy mixture model. A real dataset illustrates the usefulness of this approach.
28 A Convolutional Neural Network Combined with a Gaussian Process for Speed Prediction in Traffic Networks Yifei Zhu (The University of Sheffield)*; Peng Wang (Manchester Metropolitan University); Lyudmila Mihaylova (University of Sheffield) This paper proposes a traffic speed prediction framework combining a Convolutional Neural Network (CNN) with a Gaussian Process (GP) and is an extension of ConvNet-GP. The main focus is on spatio-temporal large scale traffic networks and on uncertainty quantification. The emphasis is on the impact on the measurement noises on the predicted traffic speeds. The Gaussian Process regression provides a variance which characterises the accuracy of the prediction. The traffic speed data is converted into a three dimensional format like images and these are inputs of the CNN-GP framework for traffic networks. The CNN-GP framework provides 18.23\% average improvement of the speed root mean square error compared with the generic CNN and gives a quantitative characterisation of the noise effects.
29 Generalization Equations for Machine Learners Based on Physical and Abstract Laws Nageswara Rao (Oak Ridge National Laboratory)* The physical and abstract laws derived from the first principles have been exploited to customize and sharpen machine learning (ML) methods and derive their generalization equations. These laws often encapsulate knowledge that complements datasets and ML models. We present a generic framework that uses these laws to provide ML codes transferable across multiple areas, including data transport infrastructures and thermal hydraulics analytics of nuclear reactors. By anchoring on datasets from these areas and the statistical generalization theory, we present a rigorous approach to co-develop ML solutions and the generalization equations that characterize them, by exploiting the structure and constraints from the physical and abstract laws. We present illustrative examples using practical problems from existing literature on performance characterization of data transport infrastructures, and power-level and sensor error estimation in nuclear reactor systems using sensor measurements of secondary and primary coolant systems, respectively.
30 Automatic Identification of the Leader in a Swarm using an Optimized Clustering and Probabilistic Approach Ajitesh Singh (University of Delaware); Panagiotis Artemiadis (University of Delaware)* Collective behavior appearing abundantly in nature at various scales, such as swarming of insects, flocking of birds or schooling of fish, has motivated an interdisciplinary research thrust towards artificial swarms thanks to its numerous advantages. Although a lot of work has been done on coming up with controllers for these multi-agent systems towards optimizing coordination and planning, the inverse problem of identifying behaviors and models from observing a swarm is still not explored enough. Efficient tools for the analysis of the complex spatial-temporal dataset an observed swarm generates, are still missing. In this paper, we propose a methodology to solve the problem of identifying the leader in a swarm governed by a leader-follower control framework using only the macroscopic view of the entire swarm. A methodology that combines clustering and probability methods is used to analyze the spatial-temporal data of a swarm of robots to narrow down on a subset of the robots that includes the leader. The clustering parameters are optimized over multiple simulated behaviors. The results show that this automated system can narrow down on a subset of agents (cluster) that includes the leader of the swarm with high accuracy. Applications of the proposed methodology include automatic identification of leader and swarm dynamics in both artificial and biological swarms, which can lead to a better understanding of collective behaviors and predictions of future behaviors.
33 CS-Fnet: A Compressive Sampling Frequency Neural Network for Simultaneous Image Compression and Recognition Rui Ma (Southern University of Science and Technology)*; Qi Hao (Southern University of Science and Technology) Data recognition using compressive measurements is desired for intelligent edge devices to save computation and communication resources. However, direct recognition of compressed image data is often difficult because the compression operation disturbs the original signal structure. In compressive sensing (CS), original signals are transformed into the frequency domain or other domains for sparse representations. This paper presents a compressive sampling frequency neural network (CS-Fnet) to achieve high computational efficiency for compressed image recognition, whose measurement matrix (MM) is automatically obtained through the CS-Fnet training. Furthermore, the MM is constructed in the form of the Kronecker product, which can reduce the number of MM parameters, and hence the CS-Fnet training can achieve much higher computational efficiency and convergence speed. The proposed method is validated using the MNIST dataset and gesture datasets. The experiment results demonstrate that the proposed CS-Fnet outperforms traditional convolution neural networks (CNNs) in terms of image recognition accuracy, and the learned MMs yield higher reconstruction accuracy than traditional MMs.
34 α-Rényi based framework for a Robust and Fault-Tolerant Localization Khoder Makkawi (CRIStAL)*; Nesrine Harbaoui (CRIStAL); Nourdine Ait-Tmazirte (University of Gustave Eiffel); Maan El Badaoui El Najjar (CRIStAL)

A robust criterion with an adaptive diagnosis for a fail-safe localization system is presented in the paper. To achieve robustness, the Minimum Error Entropy (MEE) is the criterion combined with the Unscented Information Filter (UIF) creating the MEEUIF as a robust estimator in the information theory. As known, the Unscented Transformation (UT) deals well with non-linearity problems, but the performance decreases significantly under the presence of non-Gaussian noises. The MEE overcomes this problem and shows high robustness dealing with heavy non-Gaussian noises (especially multi-Gaussian noises). For Fault Detection and Isolation (FDI), the α-Rényi divergence (α-RD) is proposed offering an adaptive diagnosis layer able to interact with the surrounding environment of the system. Then, the decision is executed based on an adaptive threshold determined through the α-Rényi criterion (α-Rc).

The proposed approach is tested and validated using real experimental data, for a multi-sensor fusion of Global Navigation Satellite System (GNSS), and Odometer (Odo) data for an autonomous vehicle localization application. The main contributions of the paper are: – The development of a robust multi-sensor fusion using the MEEUIF, – The design of an adaptive diagnosis layer based on α-Rényi divergence, and – The validation of the proposed approach using real experimental data.

35 An Inequality for Bayesian Bregman Risks With Applications in Directional Estimation Michael Fauss (Princeton University)*; Alex R Dytso (New Jersey Institute of Technology); H. Vincent Poor (Princeton University) An inequality connecting the Bayesian Bregman risk, the Kullback-Leibler divergence and distributions from the exponential family is derived. The inequality has applications in directional and robust estimation and can provide universal lower bounds on Bregman risks. Its usefulness is illustrated using the example of estimation in Poisson noise with a logarithmic cost function.
36 Target-free Extrinsic Calibration of a 3D-Lidar and an IMU Subodh Mishra (TAMU-College Station)*; Gaurav Pandey (Ford Motor Company); Srikanth Saripalli (TAMU-College Station) This work presents a novel target-free extrinsic calibration algorithm for a 3D Lidar and an IMU pair using an Extended Kalman Filter (EKF) which exploits the motion based calibration constraint for state update. The steps include, data collection by motion excitation of the Lidar Inertial Sensor suite along all degrees of freedom, determination of the inter sensor rotation by using rotational component of the aforementioned motion based calibration constraint in a least squares optimization framework, and finally, the determination of inter sensor translation using the motion based calibration constraint for state update in an Extended Kalman Filter (EKF) framework. We experimentally validate our method using data collected in our lab and open-source (https://github.com/SubMishMar/imu_Lidar_calibration) our contribution for the robotics research community.
37 Robot Localization in a Pipe Network using a Particle Filter with Error Detection and Recovery in a Hybrid Metric-Topological Space Rob Worley (The University of Sheffield)*; Sean Anderson (University of Sheffield) Estimation of a robot’s pose in the constrained environment of a buried pipe network is difficult due to uncertainty in motion, limited and unreliable sensing, and limited computational capability on board the robot. An efficient localization algorithm is described here, along with two novel improvements: the detection of mislocalization using the Kullback-Leibler divergence between the distribution of recent particle filter weights and a distribution learned from data for good algorithm performance, overcoming the problem of low information in each individual set of particle filter weights; and the capability for relocalization, using multi-hypothesis estimation to overcome the problems of limited information in sensing. The algorithm uses the low-dimensional metric-topological space of the pipe network to give efficient performance and to facilitate relocalization. Experimental results show that the localization algorithm is effective and that the presented improvements reduce the chance of unrecoverable mislocalization, therefore improving robustness to error in measurements made by the robot.
38 SemCal: Semantic LiDAR-Camera Calibration using Neural Mutual Information Estimator Peng Jiang (Texas A&M University)*; Philip Osteen (DEVCOM Army Research Laboratory (ARL)); Srikanth Saripalli (Texas A&M University) This paper proposes SemCal: an automatic, targetless, extrinsic calibration algorithm for a LiDAR and camera system using semantic information. We leverage a neural information estimator to estimate the mutual information (MI) of semantic information between sensor measurements. We use matrix exponential for transformation matrix computation and a kernel-based sampling method to sample data from camera measurement based on LiDAR projected points. We formulate the LiDAR-Camera calibration problem as a novel differentiable objective function that supports gradient-based optimization methods. We also introduce a semantic-based initial calibration method using 2D MI-based image registration and Perspective-n-Point (PnP) solver. Finally, we demonstrate the robustness of our method and quantitatively analyze the accuracy of a synthetic dataset. We also evaluate our algorithm qualitatively on an urban dataset (KITTI360) and an off-road dataset (RELLIS-3D) benchmark datasets on both ground truth labels and labels predicted by the state-of-the-art deep learning models and show improvement over recent comparable approaches.
40 Evaluation of Image Classification Networks on Impulse Sound Classification Task Ravali Nalla (Fraunhofer FKIE)*; Macarena Varela (Fraunhofer FKIE); Marc Oispuu (Fraunhofer FKIE)

Impulse sound classification has several civil and security related applications. Convolutional Neural Networks (CNNs) have been proven very effective in image classification and show promise for audio applications as well. The main aim of this paper is to see if we can use pre-defined image classification networks for impulse sound classification task and evaluate their performance on said task. We use ballistic sound dataset containing gunshots of different weapons for impulse sound classification task. For using audio segments as images, we calculate a spectrogram representation of the data and treat them as images for the training. A convolutional 2D task specific network, SimpleNet 2D and a convolutional 1D task specific network, SimpleNet 1D with similar architecture are designed and their performance is compared to study the effect of data representation on training.

From state-of-the-art of image classification networks, VGGNet, Inception v3, Inception-ResNet-v2, NASNet, ResNeXt, EfficientNet are chosen as candidate architectures for evaluation. These networks along with SimpleNet 2D are evaluated on four different test datasets and we compare the performance of the networks when trained from scratch and trained on ImageNet pre-trained weights. From these evaluations, the effect of pre-trained weights on training is observed.

41 A Question of Time: Revisiting the Use of Recursive Filtering for Temporal Calibration of Multisensor Systems Jonathan Kelly (University of Toronto)*; Christopher Grebe (University of Toronto); Matthew Giamou (University of Toronto) We examine the problem of time delay estimation, or temporal calibration, in the context of multisensor data fusion. Differences in processing intervals and other factors typically lead to a relative delay between measurement updates from disparate sensors. Correct (optimal) data fusion demands that the relative delay must either be known in advance or identified online. There have been several recent proposals in the literature to determine the delay using recursive, causal filters such as the extended Kalman filter (EKF). We carefully review this formulation and show that there are fundamental issues with the structure of the EKF (and related algorithms) when the delay is included in the filter state vector as a parameter to be estimated. These structural issues, in turn, leave recursive filters prone to bias and inconsistency. Our theoretical analysis is supported by simulation studies that demonstrate the implications in terms of filter performance; although tuning of the filter noise variances may reduce the chance of inconsistency or divergence, the underlying structural concerns remain. We offer brief suggestions for ways to maintain the computational efficiency of recursive filtering for temporal calibration while avoiding the drawbacks of the standard filtering algorithms.
43 Shape Estimation and Tracking using Spherical Double Fourier Series for Three-Dimensional Range Sensors Tim Baur (Institute of System Dynamics, HTWG Konstanz)*; Johannes Reuter (Institute of System Dynamics, HTWG Konstanz); Antonio Zea (Karlsruhe Institute of Technology); Uwe Hanebeck (Karlsruhe Institute of Technology) In this paper, a novel measurement model based on spherical double Fourier series (DFS) for estimating the three-dimensional shape concurrently to its kinematic state is introduced. Here, the shape is represented as a star-convex radial function with the function decomposed as spherical DFS. In comparison to ordinary DFS, spherical DFS do not suffer from ambiguities at the poles. Details will be given in the paper. The shape representation is integrated in a Bayesian state estimator via a measurement equation. As multilayer LiDAR sensors only generate measurements from the side facing the sensor, the shape representation is modified to enable application of shape symmetries during the estimation process. The model is analyzed in simulations and compared to a shape estimation procedure using spherical harmonics. Finally, shape estimation using spherical and ordinary DFS is compared to analyze the effect of the pole problem on extended object tracking (EOT).
44 Optimal Sensor Placement for Shooter Localization within a Surveillance Area Luisa Still (Fraunhofer FKIE)*; Marc Oispuu (Fraunhofer FKIE); Wolfgang Koch (Fraunhofer FKIE) This paper deals with optimal sensor placement for a scenario with enemy fire, where the shooter is expected to arise in a predefined surveillance area. For this purpose, an existing method that optimizes sensor positions with respect to a known and fixed shooter state is extended to a surveillance area. Here, a genetic algorithm is used to optimize the sensor positions with respect to localization accuracy given by the
associated Cramér-Rao bound. Results with different numbers and types of sensors are presented, which confirm the applicability and relevance of our method. The results show that the recommended sensor placement depends significantly on the scenario’s characteristics, such as size and location of the surveillance area or the available number of sensors.
45 Segmentation of Stereo-Camera Depth Image into Planar Regions based on Evolving Principal Component Clustering Miloš Antić (University of Ljubljana, Faculty of Electrical Engineering)*; Andrej Zdešar (Faculty of Electrical Engineering); Igor Skrjanc (n/a) In this paper we present planar segmentation of a depth image with evolving principal component clustering (EPCC). We take advantage of spatial point locality in ordered data streams to simplify development of segmentation algorithms and to speed up their operation. Experimental work has affirmed that noise modelling and its consideration in clustering process increases segmentation performance and accuracy. Evolving approach enables single pass learning and self-adaptation of linear prototypes, which are used to describe particular regions of the data. The proposed algorithm enables robust operation and adapts to data due to recursive estimation of data variance. The performance of the proposed method is demonstrated on real measurements, obtained from a depth camera.
47 Time-efficient Bayesian Inference for a (Skewed) Von Mises Distribution on the Torus in a Deep Probabilistic Programming Language Ola Rønning (University of Copenhagen)*; Christophe Ley (Ghent University); Kanti V. Mardia ( University of Leed); Thomas Hamelryck (University of Copenhagen) Probabilistic programming languages (PPLs) areat the interface between statistics and the theory of program-ming languages. PPLs formulate statistical models as stochasticprograms that enable automatic inference algorithms and opti-mization. Pyro [1] and its sibling NumPyro [2] are PPLs builton top of the deep learning frameworks PyTorch [3] and Jax [4],respectively. Both PPLs provide simple, highly similar interfacesfor inference using efficient implementations of HamiltonianMonte Carlo (HMC), the No-U-Turn Sampler (NUTS), andStochastic Variational Inference (SVI). They automaticallygenerate variational distributions from a model, automaticallyenumerate discrete variables, and support formulating deepprobabilistic models such as variational autoencoders and deepMarkov models.The Sine von Mises distribution and its skewed variant aretoroidal distributions relevant to protein bioinformatics. Theyprovide a natural way to model the dihedral angles of proteinstructures, which is important in protein structure prediction,simulation and analysis. We present efficient implementationsof the Sine von Mises distribution and its skewing in Pyroand NumPyro, and devise a simulation method that increasesefficiency with several orders of magnitude when using parallelhardware (i.e., modern CPUs, GPUs, and TPUs). We demon-strate the use of the skewed Sine von Mises distribution bymodeling dihedral angles of proteins using a Bayesian mixturemodel inferred using NUTS, exploiting NumPyro’s facilities forautomatic enumeration [5].
48 Estimation of wheel odometry model parameters with improved Gauss-Newton method Máté Fazekas (SZTAKI)*; Péter Gáspár (SZTAKI); Balázs Németh (SZTAKI) The localization of a self-driving vehicle has to be as accurate as possible for proper control and safe driving. Therefore, the GNSS, IMU, or perception-based methods should be improved, e.g. with the integration of the wheel motion. This method is robust and cost-effective, but the calibration of the model parameters behind the wheel-based odometry can be difficult. It is resulted from the nonlinear dynamics of the system and the requirement of parameter estimation with high precision, which is an open problem in the context of autonomous vehicles yet. This paper proposes an estimation architecture that applies batch mode of the Gauss-Newton estimation and compensates the noise simultaneously. With the novel architecture, the bias in the identification resulted from the state initialization uncertainty is reduced. The effectiveness of the estimation and the operation of the calibration process are illustrated through vehicle test experiments.
50 Observability-Aware Trajectory Optimization: Theory, Viability and State of the Art Christopher Grebe (University of Toronto)*; Emmett Wise (University of Toronto); Jonathan Kelly (University of Toronto) Ideally, robots should move in ways that maximize the knowledge gained about both their internal system state and the external operating environment. Trajectory design is a challenging problem that has been investigated from a variety of perspectives, ranging from information-theoretic analyses to leaning-based approaches. Recently, observability-based metrics have been proposed to find trajectories that enable rapid and accurate state and parameter estimation. The viability and efficacy of these methods is not yet well understood in the literature. In this paper, we compare two state-of-the-art methods for observability-aware trajectory optimization and seek to add important theoretical clarifications and valuable discussion about their overall effectiveness. For evaluation, we examine the representative task of sensor-to-sensor extrinsic self-calibration using a realistic physics simulator. We also study the sensitivity of these algorithms to changes in the information content of the exteroceptive sensor measurements.
52 Towards Robust VSLAM in Dynamic Environments: A Light Field Approach Pushyami Kaveti (Northeastern University)*; Jagatpreet Singh Nir (Northeastern University); Hanumant Singh (Northeastern University) There is a general expectation that robots should operate in urban environments often consisting of potentially dynamic entities like people, automobiles etc. Dynamic objects pose challenges to visual SLAM algorithms by introducing errors into the front-end. This paper presents a Light Field powered SLAM Frontend which is robust to dynamic environments. A Light Field captures a bundle of light rays emerging from a single point in space, allowing us to see through dynamic objects occluding the static background via Synthetic Aperture Imaging(SAI). We detect apriori dynamic objects using semantic segmentation and perform semantic guided SAI on the Light Field acquired from a linear camera array. The combined use of SAI with semantic segmentation results in a significant reduction of the dynamic content and also adds valuable observations from the occluded static parts of the world. The GPU implementation of the algorithm facilitates running at a speed of $\sim$6 fps. We demonstrate considerable improvement in robustness and accuracy of pose estimation in dynamic environments by comparing it with state-of-the-art VSLAM algorithms.Link to Video demo: \url{https://rb.gy/mwt5tv}
53 Object Detection and Mapping with Bounding Box Constraints Benchun Zhou (KIT)*; Aibo Wang (KIT); Jan-Felix Klein (KIT) In this paper, we present a three-dimensional object detection method for a single image and an object-based localization and mapping system. For 3D object detection, we firstly generate high-quality cuboid candidates by sampling object rotation and dimension. Then, the translation of each candidate is estimated in a closed form solution with camera projection function and bounding box constraints. Finally, all candidates are projected into the image, scored and selected based on the alignment with detected lines. To overcome object detection accuracy issues, the results are improved by multi-view optimization. Besides, objects can provide geometry constraints and semantic information to improve camera pose estimation and monocular drift. A point-object SLAM system is formulated to jointly optimize the poses of camera, objects and points. We evaluate our object detection method on objects from the KITTI, the SUN RGB-D and a self collected dataset. The results show that our method outperforms existing approaches. The point-cuboid SLAM experiments on the TUM RGB-D, ICL-NUIM and our self collected dataset show that our algorithm can improve both camera localization accuracy and 3D object detection accuracy.
55 Latency Analysis of ROS2 Multi-Node Systems Tobias Kronauer (Barkhausen Institut)*; Joshwa Pohlmann (Barkhausen Institut); Maximilian Matthe (Barkhausen Institut); Till Smejkal (TU Dresden); Gerhard Fettweis (Barkhausen Institut)

The Robot Operating System 2 (ROS2) targets distributed real-time systems and is widely used in the robotics community. Especially in these systems, latency in data processing and communication can lead to instabilities. Though being highly configurable with respect to latency, ROS2 is often used with its default

In this paper, we investigate the end-to-end latency of ROS2 for distributed systems with default settings and different Data Distribution Service (DDS) middlewares. In addition, we profile the ROS2 stack and point out latency bottlenecks. Our findings indicate that end-to-end latency strongly depends on the used DDS middleware. Moreover, we show that ROS2 can lead to 50% latency overhead compared to using low-level DDS communications. Our results imply guidelines for designing distributed ROS2 architectures and indicate possibilities for reducing the ROS2 overhead.

56 UKF Parameter Tuning for Local Variation Smoothing Kristin Nielsen (Linköping University)*; Caroline Svahn (Linköping University); Hector Rodriguez-Deniz (Linköping University); Gustaf Hendeby (Linköping University) The unscented Kalman filter (UKF) is a method to solve nonlinear dynamic filtering problems, which internally uses the unscented transform (UT). The behavior of the UT is controlled by design parameters, seldom changed from the values suggested in early UT/UKF publications. Despite the knowledge that the UKF can perform poorly when the parameters are improperly chosen, there exist no wide spread intuitive guidelines for how to tune them. With an application relevant example, this paper shows that standard parameter values can be far from optimal. By analyzing how each parameter affects the resulting UT estimate, guidelines for how the parameter values should be chosen are developed. The guidelines are verified both in simulations and on real data collected in an underground mine. A strategy to automatically tune the parameters in a state estimation setting is presented, resulting in parameter values inline with developed guidelines.
59 Gaussian Mixture Estimation from Weighted Samples Daniel Frisch (Karlsruhe Institute of Technology)*; Uwe Hanebeck (Karlsruhe Institute of Technology) We consider estimating the parameters of a Gaussian mixture density with a given number of components best representing a given set of weighted samples. We adopt a density interpretation of the samples by viewing them as a discrete Dirac mixture density over a continuous domain with weighted components. Hence, Gaussian mixture fitting is viewed as density re-approximation. In order to speed up computation, an expectation–maximization method is proposed that properly considers not only the sample locations, but also the corresponding weights. It is shown that methods from literature do not treat the weights correctly, resulting in wrong estimates. This is demonstrated with simple counterexamples. The proposed method works in any number of dimensions with the same computational load as standard Gaussian mixture estimators for unweighted samples.
63 Learning and Exploiting Partial Knowledge in Distributed Estimation Susanne Radtke (Karlsruhe Institute of Technology)*; Jiri Ajgl (University of West Bohemia); Ondrej Straka (University of West Bohemia); Uwe Hanebeck (Karlsruhe Institute of Technology) In distributed estimation, several sensor nodes provide estimates of the same underlying dynamic process. These estimates are correlated but due to local processing, the correlations are only partially known or even unknown. For a consistent fusion of the local estimates, the correlation needs to be properly treated. Many methods provide consistent but overly conservative fusion results. In this paper, we propose to learn partial knowledge about the correlation in the form of correlation sets and exploit this knowledge to provide less conservative estimates. We use a simple numerical example to demonstrate the advantages of the proposed approach in terms of quality and consistency and how the quality of the fused estimate increases with time.
64 Directional DD-classifiers under non-rotational symmetry Houyem Demni (Universtity of Cassino and Southern Lazio)*; Giovanni Camillo Porzio (Universtity of Cassino and Southern Lazio) A directional random variable is rotationally symmetric around a location parameter if its distribution only depends on the angle (colatitude in case of a sphere) between the value the variable can take and the location parameter itself. This is clearly an oversimplified model in many practical applications. On the other hand, the performance of depthbased classifiers for directional data has been widely studied only under the case of rotational symmetry of the underlying class distributions. For this reason, this work aims at evaluating the efficacy of some of these classifiers under non-rotational symmetry. Particularly, the DD-classifiers exploiting the linear, quadratic, and KNN discriminant rules when associated with angular distance-based depths are examined. Their performances under Kent distributions are investigated by means of a simulation study. As a benchmark, the directional Bayes rule is considered. In passing, these classifiers are also reviewed, noting that the way they work within the directional data domain has been given a bit for granted within the literature.
65 Incorporating Range Rate Measurements in EKF-based Elliptical Extended Object Tracking Kolja Thormann (Universtiy of Goettingen)*; Marcus Baum (University of Goettingen) This work deals with the tracking of an elliptical extended object using radar measurements. We model the ellipse using the orientation and semi-axes as individual state parameters. The radar measurements typically consist of range, bearing, and range rate with individual noise. To minimize non-linearities in the measurement equation, we utilize the transformation of the measurement noise from polar to Cartesian coordinates along with the removal of that transformation’s bias and we calculate a pseudo range rate by multiplying range and range rate measurement. Using the resulting measurement vector, we define a measurement equation based on multiplicative noise and apply linearization techniques to get a closed-form solution for the update using Kalman filter formulas. We evaluate our approach in simulations, comparing it to the original MEM-EKF* and an approach using linearization to incorporate Doppler range rate into the random matrix framework.
66 Kalman Filtered Compressive Sensing Using Pseudo-Measurements Haibin Zhao (KIT, TecO)*; Christopher Funk (OVGU); Benjamin Noack (OVGU); Uwe Hanebeck (Karlsruhe Institute of Technology); Michael Beigl (Karlsruhe Institute of Technology) In this paper, we combine the Kalman filter and compressive sensing using pseudo-measurements in order to reduce the number of measurements usually required by the Kalman filter. To overcome the non-sparsity of the measurement vectors, we make use of the change of their coefficients when represented in a certain basis, reduce the dimensionality of the coefficients, and learn a sparse basis for the measurement vectors. We further improve our proposed method by introducing dynamic weighting of the pseudo-measurements, by aiding compressive measurement reconstruction with Kalman filter estimates and by employing iterative versions of this process. Simulations show that our approach achieves a 40% improvement with respect to the mean-square error compared to the traditional Kalman filter with the same number of measurements. Our approach yields better results when the measurement noise is relatively large compared to the system noise, and it significantly improves the accuracy of state estimation in sensor networks with low sensor precision.
68 Skeleton-based Fall Events Classification with Data Fusion Leiyu Xie (Newcastle University)*; Yuxing Yang (Newcastle University); Zeyu Fu (University of Oxford); Syed Mohsen Naqvi (Newcastle University) Human fall detection aims to classify falls and normal activities. It can improve the speed of rescue for the aging people after a fall occurs, it can also efficiently prevent aging people from suffering secondary injuries due to untimely or inaccurate fall detection. This technology is widely used in hospitals, smart homes and nursing homes. The challenges are that the injured parts of aging people are various due to different types of fall events, such as falling sideways and falling backwards. In this paper, we proposed a data fusion method which combined the skeleton keypoints captured from RGB images into fused keypoints to improve the performance of fall events classification. Meanwhile, the impact on fall detection results due to data missing caused by occlusion and private-privacy situations is discussed through ablation study. The experimental results confirm that the proposed framework are outperformed than those in state-of-the-art with reduced computational cost.
69 Conditional Densities and Likelihoods for Hypertoroidal Densities Based on Trigonometric Polynomials Florian Pfaff (Karlsruhe Institute of Technology)*; Kailai Li (Karlsruhe Institute of Technology); Uwe Hanebeck (Karlsruhe Institute of Technology) Recently, trigonometric polynomials have been used to approximate densities or their square roots in the context of Bayesian estimation. Trigonometric polynomials were also used to interpolate function values on grids on hypertori. In this paper, we derive formulae for conditional densities and likelihoods for multivariate densities parameterized by grid values or Fourier coefficients. Efficient formulae are proposed for both representations that involve no more than O(n log n) operations. The conditional densities can be described using a single parameter vector. For the likelihoods, formulae are given that allow for a precise evaluation using two parameter vectors. Furthermore, formulae involving only a single parameter vector are provided for approximations of the likelihoods.
70 Mono-Vision based Moving Object Detection using Semantic-Guided RANSAC songming CHEN (Ecole centrale de Nantes)*; Haixin Sun (Ecole centrale de Nantes); Vincent Fremont (Centrale Nantes – LS2N UMR 6004)

This paper proposes a novel two-stage approach for detecting moving objects with a non-stationary monocular camera mounted on a vehicle. We formulate an innovative method called semantic-guided random sample consensus (Semantic-Guided RANSAC) to detect moving objects by semantic-geometric information fusion and integration. Firstly, semantic constraints from deep learning architecture (YOLO v4) are applied to predict the objects’ location in the image frame. The fundamental matrix is then estimated robustly from two views through the sparse optical flow tracking with the help of semantic prior. Semantic-guided RANSAC is used to reject instance-level outliers which are actually moving objects based on the epipolar geometry and flow vector bound constraints. Experimental results on KITTI dataset reflect the effectiveness of our approach to identify moving objects in complex urban traffic scenes with the average precision above 0.82 for 4 sequences in the City category.