Climate change is intensifying precipitation events, leading to increased flooding and landslides, particularly in urban areas. Traditional Numerical Weather Prediction (NWP) models are essential for forecasting these events, but require extensive computational resources for high spatio-temporal resolution. Machine Learning (ML) approaches offer a more computationally efficient alternative, with fast prediction phases suitable for less powerful resources, while intensive training can be done asynchronously. We propose a deep learning architecture that incorporates an information data fusion module prior to a CNN-based decision module to reduce the input size while maintaining prediction accuracy. By testing the application of two different subspace methods, Principal Component Analysis decomposition (PCA) and Maximum pixel map, we reduce the number of input channels needed for the decision module. Experiments on the rainfall dataset compare the performance of the decision module with and without the fusion module. The experiments conclude that while the baseline original SmaAt-UNet remains the most reliable, the Fusion Module Max Pixel Map provides a substantial input compression while maintaining performance. This could enable the execution of forecasts on devices with reduced processing capabilities, such as mobile devices.
The poses derived from automatic point cloud registration between stationary laser scans that occur naturally in robotic multi-sensor systems operating in stop-and-go mode have the potential to greatly improve and aid localization and trajectory estimation. The uncertainty assessment of these registered poses is crucial for its correct system integration and utilization, e.g., for mobile mapping applications. However, obtaining ground truth data with sufficient accuracy for both the position and attitude remains challenging. In this paper, we present a novel measurement setup and a data fusion method based on an iterative weighted least squares adjustment to assess the quality of such derived relative poses. The approach is tested and demonstrated in an outdoor experiment covering a driving area of approximately 65 x 7 m. A total station and a laser tracker were used to realize the ground truth. The latter was mounted on a Husky A200 UGV together with a Riegl VZ-600i laser scanner. The realized ground truth enables the determination of the relative poses with respect to a superordinate system with 0.3 mm and 0.04 mrad on average. The relative poses of the automatic registration differ from the reference poses up to a maximum of 3 mm and 0.76 mrad.
In autonomous robotics, a critical challenge lies in developing robust solutions for Active Collaborative SLAM, wherein multiple robots collaboratively explore and map an unknown environment while intelligently coordinating their movements and sensor data acquisitions. In this article, we present an efficient centralized frontier sharing approach that maximizes exploration by taking into account information gain in the merged map, distance, and reward computation among frontier candidates and encourages the spread of agents into the environment. Eventually, our method efficiently spreads the robots for maximum exploration while keeping SLAM uncertainty low. Additionally, we also present two coordination approaches, synchronous and asynchronous to prioritize robot goal assignments by the central server. The proposed method is implemented in ROS and evaluated through simulation and experiments on publicly available datasets and similar methods, rendering promising results.
The Fisher information number (FIN) has previously been proposed as a regularizer to fit a probability density function to a set of constraints. Especially for mixture densities, this is not straightforward and often a reformulation based on square root densities is used. As it is generally much harder to derive the square root of a mixture than squaring it, this only allows for constraints that can be expressed through the root density's parameters. An important case not covered by this are constraints on individual components of a mixture. This paper proposes three methods to approximate the FIN of mixture models: Gauss-Hermite quadrature, polynomial approximation of the square root function, and direct approximation of the square root density of a pdf. This allows using the FIN for smooth density estimation in situations existing methods cannot handle. The three methods are applied to the problem of kernel density estimation with Gaussian kernels and the results are compared.
Magnetic field vector measurement model derived from a simple geometric coil model has been widely reported. These field vector measurements can be used to estimate the pose of a vehicle at a high level of accuracy in real-time. In our previous work, we introduced a magnetic positioning system (MPS) using these magnetic field vector measurements and performed 2D pose estimation. In this paper, we first introduce the initial calibration of the magnetic field considering both the magnetic field measurements and theoretical models. Then, estimation of position, velocity, and attitude is performed by multi sensor fusion between calibrated MPS, odometer and INS. This is implemented in a tightly-coupled(TC) extended Kalman filter(EKF) structure. The results demonstrated precise three- dimensional navigation performance and the benefits of the tightly coupled structure, resulting in accurate navigation and robustness.
In this paper, we propose a novel Gaussian Assumed Density Filter (GADF) for high-quality state estimation of nonlinear dynamic systems. Our approach focuses on the measurement update, utilizing a non-Gaussian local approximation of the true joint measurement/prior state density. This is achieved through the sequential use of two Inverse Gaussian Processes (IGPs): the first IGP interpolates the means, and based on these results, the second IGP is trained to interpolate the covariances. Together, they fully characterize the conditional Gaussian densities of the hidden state on concrete measurements. Consequently, our method does not require the second Gaussian assumption for the joint density anymore, thereby enhancing filter performance. Moreover, our approach eliminates the need for an explicit likelihood function within the filter step, making it a higher-quality plug-in replacement for the commonly used Linear Regression Kalman Filter (LRKF).
The representation of a discrete state space in a quantum machine can be very effective due to the superposition effect, where $n$ qubits represent 2 n states. To the authors' knowledge, the Bayesian update step using a quantum computing regime has been unsolved up to date. In this paper, a quantum realization of the Daum-Huang log-homotopy flow filter is presented as a solution of the filtering step in a Bayesian recursion for state estimation. The numerical example for a low-dimensional problem demonstrates the feasibility of the presented approach.
Today's autonomous vehicles rely on a multitude of sensors to perceive their environment. To improve the perception or create redundancy, the sensor's alignment relative to each other must be known. With Multi-LiCa, we present a novel approach for the alignment, e.g. calibration. We present an automatic motion- and targetless approach for the extrinsic multi LiDAR-to-LiDAR calibration without the need for additional sensor modalities or an initial transformation input. We propose a two-step process with feature-based matching for the coarse alignment and a GICP-based fine registration in combination with a cost-based matching strategy. Our approach can be applied to any number of sensors and positions if there is a partial overlap between the field of view of single sensors. We show that our pipeline is better generalized to different sensor setups and scenarios and is on par or better in calibration accuracy than existing approaches. The presented framework is integrated in ROS 2 but can also be used as a standalone application. To build upon our work, our source code is available at https://github.com/TUMFTM/Multi_LiCa.
The introduction of autonomous vehicles (AVs) presents a novel approach to regulating and optimising traffic flow through the automated control of AVs. In this context, the AV is defined as the actuator and an optimal control policy is desired to make control decisions. Deep Reinforcement Learning (DRL) is a novel method which aims to maximize the cumulative rewards given by the predefined reward function by making sequential decisions in a stochastic environment. In light of the above, we propose a DRL-based vehicular control method to train an optimal policy for the control of AV in a model-free fashion, and consequently improve the traffic efficiency with the obtained control policy. A single-lane circular road environment with both AV and human-driven vehicles is selected to serve as the mixed autonomy traffic system in the Simulation of Urban MObility (SUMO) [1] traffic simulator, and the Proximal Policy Optimization (PPO) algorithm is applied for the policy improvement. Simulation results demonstrate that our strategy is effective in mitigating the unstable stop-and-go waves, increasing 67.7% of the average driving speed and reducing 19.3% of the average energy consumption in a closed-ring road environment.
Data fusion is the process of integrating data from multiple sources to produce more accurate and reliable information. It is often the case that data are subject to latent confounding and measurement error in real-world scenarios. In this paper, we evaluate fusion strategies based on different levels of contained causal knowledge to solve quality prediction under varied conditions of latent confounding and measurement error. We show that the machine learning-based fusion strategy achieves the best prediction quality when data are independent and identically distributed (i.i.d.). However, in the presence of latent confounding, the causality-based fusion strategy makes prediction models more robust against severe distribution shifts. Moreover, the out-of-distribution (OOD) generalizability of prediction models is also affected by measurement error in the data. If causal knowledge needs to be inferred from data by applying causal discovery methods, we demonstrate that measurement error can adversely impair causal discovery. We advocate that caution needs to be exercised when using standard causal discovery methods if the circumstances under which the data were generated are unknown.
Reliably monitoring the environment with distributed sensors is a necessity for many modern automation tasks such as automated driving. However, the growing demand for communications resources can hardly be fulfilled in the future without a paradigm shift in resource utilization. One way to leverage the burden on the communications system is to transmit data in an event-based fashion rather than periodically at a high rate. Several event-based triggers and estimators have been proposed in the past. Unfortunately, the event-based schemes are often sensitive to imperfections in the communications system such as packet losses. To ensure reliable estimates under packet losses, a new stochastic event-based scheme is proposed that uses the transmission probability of the trigger as an additional periodic information source in the estimator on the receiver side. The effectiveness of the approach is evaluated in simulation using different packet loss models.
This paper addresses a risk-sensitive remote estimation problem for cyber-physical systems (CPSs) where the accurate model of a dynamic system is not completely known or may differ from the assumed model. In CPSs, sensors and the monitoring control center are remotely located. Sensors transmit the measurements via unreliable wireless communication channels that are vulnerable to cyber-attacks. Specifically, attackers can inject false data to alter the measurements in the communication channel or attack sensors. To tackle this, we design a risk-sensitive filtering algorithm to operate under false data injection attacks. The proposed estimator aims to minimize the risk-sensitive error criterion, defined as the expectation of the accumulated exponential quadratic error. Simulation results demonstrate the effectiveness of the proposed algorithm.
A simple and real-time 2D single-object visual tracker is derived based on first-principle modeling in 3D, which was introduced in previous work. An inverse of the nonlinear perspective projection is followed by a simple approximation assuming that the pedestrian moves with a fixed depth in front of the camera. The resulting 2D tracking algorithm appears to have a similar form as the state-of-the-art single-object tracker from BoT-SORT, is easy to implement, and its parameters have convenient meaning. Its performance is assessed statistically using the publicly available MOT17 dataset.
Linear fusion of estimates under the condition of no knowledge of correlation of estimation errors has reached maturity. On the other hand, various cases of partial knowledge are still active research areas. A frequent motivation is to deal with “common information” or “common noise”, whatever it means. A fusion rule for a strict meaning of the former expression has already been elaborated. Despite the dual relationship, a strict meaning of the latter one has not been considered so far. The paper focuses on this area. The assumption of unknown “common noise” is formulated first, analysis of theoretical properties and illustrations follow. Although the results are disappointing from the perspective of a single upper bound of mean square error matrices, the partial knowledge demonstrates improvement over no knowledge in suboptimal cases and from the perspective of families of upper bounds.
Ensuring operational safety and compliance through effective detection of abnormal drone behaviors in urban environments is a significant challenge. This study enhances anomaly detection methodologies using Dynamic Bayesian Networks (DBNs) to model complex, temporal behaviors, addressing the lack of real observational data. We extend anomaly detection to track behaviors over time, offering a nuanced understanding of drone dynamics. A novel framework is introduced, leveraging simulated environments to generate rich datasets of normal and abnormal behaviors, integrated with data fusion techniques like the Joint Probabilistic Data Association Filter (JPDAF) to amalgamate diverse sensor inputs. Additionally, the Belief Function Machine (BFM) is incorporated to handle operational uncertainties and fuse incomplete or conflicting information, further improving the model's efficiency and realtime applicability in urban airspace management. Through rigorous experimentation, our research bridges the gap in drone anomaly detection and sets a benchmark for autonomous aerial vehicle monitoring. This work highlights the importance of advanced computational models and simulation-based testing in overcoming real-world data limitations, advancing drone safety and regulatory compliance in urban settings.
Bayesian neural networks (BNNs) offer an elegant and promising approach to quantifying the uncertainty of neural network predictions by providing predictive distributions. Although the potential of BNNs is considerable, established BNN training methods often result in inaccurate uncertainty estimation and local differences in quality depending on the considered input space region. To assess the efficacy of Bayesian models such as BNNs and gain insights into their predictive capabilities in distinct input space regions, we introduce a novel methodology that utilizes ball trees as a space partitioning data structure. Our approach enables the assessment of the predictive quality within specific regions of the input space across multiple scales in the input space, utilizing all nodes provided by the ball tree structure. Furthermore, our method allows the combination of results across different scales.
Energy consumption is a critical factor that nega-tively impacts the environment. Sustainable production is essen-tial for addressing the climate crisis, as low-emission manufacturing can both reduce costs and minimize environmental impact. Energy-efficient CNC machine tools significantly contribute to achieving ambitious environmental objectives. In recent years, numerous studies have focused on low-energy consumption production, analyzing factors that contribute to sustainable manufacturing. When using the analytical or empirical model, factors and corrections might be omitted. With advancements in machine learning and the increasing availability of large datasets, models are being developed to predict energy consumption with high accuracy. However, these models often overlook the importance of features that contribute to a transparent prediction process and their influence on the results. In our paper, a LSTM model is initially utilized to predict the energy consumption of CNC machines. Following this, a method is devised to rank the features based on their predictive power, considering temporal variations. We show that some of the features ranked in the top positions agree with independent literature findings, while others are new and demand further investigation.
This paper presents a comparison between the recently proposed KalmanNet for dynamic state estimation with unknown measurement and dynamic noise covariance matrices, and a classic approach to solve this problem. Given known transition and measurement functions and a training data set that consists of sequences of ground truth states and the associated measurements, KalmanNet learns the parameters of a network that aims to compute the Kalman gain. The classic approach we consider is to estimate the noise covariance matrices via maximum likelihood estimation (MLE) during training. Then, a Kalman filter with the estimated covariance matrices is used during testing (Kalman-MLE). The benefits of Kalman-MLE versus KalmanNet are shown via experiments in two linear-Gaussian systems, and a non-linear system.
Registration of point clouds is a fundamental task in robotic SLAM pipelines. Typically this task is performed only on point clouds of the same sensor or at least the same sensing modality. However, robots designed for challenging environments are often equipped with redundant sensors for the same task where some sensors are more accurate and others are more robust against disturbing environmental conditions. Being able to register the data across the modalities is an important step to more fault-tolerant localization and mapping. We therefore propose a learning framework, which describes the points in the point cloud invariant of their modality. This description is then used in a transformer-like model to find point matches for the registration process. We demonstrate our results using a scanning lidar and radar sensor on our own and publicly available datasets.
In Random Finite Set based multi-sensor multi-object tracking, the NP-hard measurement-to-track assignment problem is a key challenge. One approach to address this challenge involves executing computationally simpler single-sensor updates based on a common prediction, followed by the fusion of these updates. This strategy is used by the already proposed Fast Product Multi-Sensor Labeled Multi-Bernoulli filter, which still poses computational challenges in its existing formulation. This paper introduces an efficient implementation of the Fast Product Multi-Sensor Labeled Multi-Bernoulli filter by improving the efficiency of the individual single-sensor updates and the fusion of the resulting single-sensor posterior densities. Two different approaches are presented for each part, including the new GeneralizedKBestSelection algorithm, which solves a $k$ shortest path problem on highly structured graphs. Our approach is evaluated on simulations, and the results are compared with an Iterated Corrector implementation of the Labeled Multi-Bernoulli filter using comparable simplifications.
