Fraunhofer-Institut für Optronik, Systemtechnik und Bildauswertung IOSB

Filters

3
Reset filters

Settings
Now showing 1 - 10 of 2787
  • Publication
    A Cross Branch Fusion-Based Contrastive Learning Framework for Point Cloud Self-supervised Learning
    ( 2024)
    Wu, Chengzhi
    ;
    Huang, Qianliang
    ;
    Jin, Xingkun
    ;
    Pfrommer, Julius
    ;
    Beyerer, Jürgen
    Contrastive learning is an essential method in self-supervised learning. It primarily employs a multi-branch strategy to compare latent representations obtained from different branches and train the encoder. In the case of multi-modal input, diverse modalities of the same object are fed into distinct branches. When using single-modal data, the same input undergoes various augmentations before being fed into different branches. However, all existing contrastive learning frameworks have so far only performed contrastive operations on the learned features at the final loss end, with no information exchange between different branches prior to this stage. In this paper, for point cloud unsupervised learning without the use of extra training data, we propose a Contrastive Cross-branch Attention-based framework for Point cloud data (termed PoCCA), to learn rich 3D point cloud representations. By introducing sub-branches, PoCCA allows information exchange between different branches before the loss end. Experimental results demonstrate that in the case of using no extra training data, the representations learned with our self-supervised model achieve state-of-the-art performances when used for downstream tasks on point clouds.
  • Publication
    Utilizing Dataset Affinity Prediction in Object Detection to Assess Training Data
    ( 2024)
    Becker, Stefan
    ;
    Bayer, Jens
    ;
    Hug, Ronny
    ;
    Hübner, Wolfgang
    ;
    Arens, Michael
    Data pooling offers various advantages, such as increasing the sample size, improving generalization, reducing sampling bias, and addressing data sparsity and quality, but it is not straightforward and may even be counterproductive. Assessing the effectiveness of pooling datasets in a principled manner is challenging due to the difficulty in estimating the overall information content of individual datasets. Towards this end, we propose incorporating a data source prediction module into standard object detection pipelines. The module runs with minimal overhead during inference time, providing additional information about the data source assigned to individual detections. We show the benefits of the so-called dataset affinity score by automatically selecting samples from a heterogeneous pool of vehicle datasets. The results show that object detectors can be trained on a significantly sparser set of training samples without losing detection accuracy. T
  • Publication
    Activities that Correlate with Motion Sickness in Driving Cars – An International Online Survey
    ( 2024)
    Diederichs, Frederik
    ;
    Herrmanns, Amina
    ;
    Lerch, David
    ;
    Zhong, Zeyun
    ;
    Daniela Piechnik
    ;
    Mathis, Lesley-Ann
    ;
    Xian, Boyu
    ;
    Vaupel, Nicklas Jakob Elia
    ;
    Vijayakumar, Ajona
    ;
    Cabaroglu, Canmert
    ;
    Rausch, Jessica
    Up to 2 out of 3 passengers suffer from motion sickness, caused by non-driving related activities. Occupant monitoring systems detect such activities via cameras in the vehicle interior and hence can be used to warn passengers or to assist them. An international online survey in Germany, USA, China, India, Turkey and Mexico was conducted in order to identify activities that correlate with motion sickness. The results identify reading, using a device, watching a movie and turning in the seat to be the most relevant activities for occupant monitoring systems to detect and hence for motion sickness assistance systems to address.
  • Publication
    Requirements Analysis for the Evaluation of Automated Security Risk Assessments
    ( 2024)
    Ehrlich, Marco
    ;
    Lukas, Georg
    ;
    Trsek, Henning
    ;
    Jasperneite, Jürgen
    ;
    Kastner, Wolfgang
    ;
    Diedrich, Christian
    The overall Industry 4.0 developments and the highly dynamic threat landscape enhance the need for continuous security engineering of industrial components, modules, and systems. Security risk assessments play a major role to ensure a secure operation of Industrial Automation and Control Systems (IACSs) but are mostly neglected due to missing resources and a lack of human experts for the sophisticated manual tasks. Therefore, a method for information and process modelling regarding the automation of security risk assessments has been previously designed, but not yet evaluated. This work in progress begins the evaluation of the automated security risk assessment concept by investigating the related work and identifying the main deficits. The results include a requirements analysis for the verification and an outlook towards future evaluation aspects.
  • Publication
    Sensitivity Analysis and Extended Evaluation of the Two-Level Stochastic Model for the Lateral Movement of Vehicles within their Lane
    ( 2024)
    Neis, Nicole
    ;
    Beyerer, Jürgen
    The range of vision of vehicle sensors used by auto-mated driving functions is considerably influenced by the lateral movement of vehicles within their lane, both of the recording vehicle and of vehicles around it. To cover this phenomenon in simulations for virtual validation, a two-level stochastic model for the lateral movement of vehicles within their lane under homogeneous traffic conditions has recently been proposed by Neis et al. [1]. It consists of the superposition of a Markov model for the discrete and systematic coarse movement, and a noise model for the stochastically independent fine movement. The general importance of the parameter choice on the performance of the two-level model has already been noted when the model was introduced. More details are added by the present work performing a sensitivity analysis on the model's parameters. Moreover, the evaluation of the model is extended by assessing its capability to describe the lateral speed. The work deepens the understanding of the model's way of working and its sensitivity to the parameters, and motivates the parameter choice in previous work. Besides this, potential for improvement of the model beyond what has already been discussed in the introductory paper is identified based on the extended investigation.
  • Publication
    Enhancing Skeleton-Based Action Recognition in Real-World Scenarios Through Realistic Data Augmentation
    ( 2024)
    Cormier, Mickael
    ;
    Schmid, Yannik
    ;
    Beyerer, Jürgen
    Skeleton-based action recognition is a prominent research area that provides a concise representation of human motion. However, real-world scenarios pose challenges to the reliability of human pose estimation, which is fundamental to such recognition. The existing literature mainly focuses on laboratory experiments with near-perfect skeletons, and fails to address the complexities of the real world. To address this, we propose simple yet highly effective data augmentation techniques based on the observation of erroneous human pose estimation, which enhance state-of-the-art methods for real-world skeleton-based action recognition. These techniques yield significant improvements (up to +4.63 accuracy) on the widely used UAV Human Dataset, a benchmark for evaluating real-world action recognition. Experimental results demonstrate the effectiveness of our augmentation techniques in compensating for erroneous and noisy pose estimation, leading to significant improvements in action recognition accuracy. By bridging the gap between laboratory experiments and real-world scenarios, our work paves the way for more reliable and practical skeleton-based action recognition systems. To facilitate reproducibility and further development, the Skelbumentations library is released at https://github.com/MickaelCormier/Skelbumentations. This library provides the code implementation of our augmentation techniques, enabling researchers and practitioners to easily augment skeleton sequences and improve the performance of skeleton-based action recognition models in real-world applications.
  • Publication
    Causal Structure Learning Using PCMCI+ and Path Constraints from Wavelet-Based Soft Interventions
    ( 2024)
    Rehak, Josephine
    ;
    Falkenstein, Alexander
    ;
    Beyerer, Jürgen
    The discovery of causal relations via interventions has proven to be simple when only one observed variable is affected or unaffected. However, in a multivariate setting, it is likely that more than one variable is affected by an intervention and thus drawing conclusions about the causal relations becomes more difficult as the gained information is ambiguous. To deal with this, we introduce a novel definition of path constraints and to obtain them, we came up with a novel approach for wavelet-based interventions. We demonstrate our approach on a combustion engine simulation, where we injected wavelets of our choice in an actuated variable and tried to rediscover them in the other, observed variables to gain path constraints. Subsequently, we demonstrate how to use these constraints to optimize the results of the established PCMCI+ algorithm.
  • Publication
    Comparison for thermal imager performance assessment: TOD classifier versus YOLO-based models for object detection
    ( 2024)
    Wegner, Daniel
    ;
    Keßler, Stefan
    Models for triangle orientation discrimination (TOD) have been proposed for performance evaluation of thermal imaging devices. For thermal imager assessment, human visual systems for TOD have been modeled and rigorously validated for a wide variety of image distortions through observer studies. As the conduct of observer trials is time-consuming and costly, also AI-based TOD models for imager assessment have been presented. Recently, camera systems with embedded automatic target recognition (ATR) are becoming increasingly important. So far it is an open question if the simple TOD task, as a classification problem with 4 classes, is suitable for providing similar evaluations and rankings for these thermal imaging devices as algorithms for more complex and slower tasks like object detection, e.g. for ATR. A widely used framework for object detection is “You Only Look Once” (YOLO). In this work, performance assessments for TOD models and YOLO-based models are compared. Known image databases as well as synthetic images with triangles and natural backgrounds are degraded according to a unified device description with blur and image noise. The blur caused by optical diffraction and detector footprint is varied by multiple aperture diameters and detector sizes through the application of modulation transfer functions, while the image noise is varied by multiple noise error levels as Gaussian sensor noise. The TOD models are evaluated for the degraded images with triangles, while the YOLO models are applied to the degraded variants of the image databases. For different degradation parameters, the model precisions of the TOD models are compared to figures of merit of the YOLO models such as the mean average precision (mAP). Statistical uncertainties of the performance ranking for different degradation parameters of cameras and both TOD and YOLO models are investigated.
  • Publication
    Remote Quantum Ghost Imaging
    ( 2024)
    Pitsch, Carsten
    ;
    Walter, Dominik
    ;
    Bürsing, Helge
    ;
    Eichhorn, Marc
    Quantum Ghost Imaging (QGI) is a scheme using entangled pairs of photons (signaland idler) in order to perform imaging with both single photons and with only a single-element detectorin the spectrum of interest. It utilises the temporal coincidence of the photons to identify associated pairs, while their spatial correlation allows to obtain image information from of the idler photon from the measurement of the signal photon.It is especially useful when using non-degenerate photon pairs, allowing to keep the signal photon in the silicon detection window, while the interacting wavelength can be freely chosen.However, current schemes are limited, as they rely on time-gating and heralding. Recent advances in single photon avalanche diodes (SPADs) allow the design of new single photon cameras, which can be outfitted with dedicated in-pixel timing circuitry. This allows to register single photons in both time and space. These detectors allowed us to design a new scheme for QGI, in which the coincidence is evaluated after the measurement. It also allows us to perform depth-resolved 3D imaging based on the time-of-flight of photons,first results of which are presented here.
  • Publication
    Unsupervised 3D Skeleton-Based Action Recognition using Cross-Attention with Conditioned Generation Capabilities
    ( 2024)
    Lerch, David
    ;
    Zhong, Zeyun
    ;
    Martin, Manuel
    ;
    Voit, Michael
    ;
    Beyerer, Jürgen
    Human action recognition plays a pivotal role in various real-world applications, including surveillance systems, robotics, and occupant monitoring in the car interior. With such a diverse range of domains, the demand for generalization becomes increasingly crucial. In this work, we propose a cross-attention-based encoder-decoder approach for unsupervised 3D skeleton-based action recognition. Specifically, our model takes a skeleton sequence as input for the encoder and further applies masking and noise to the original sequence for the decoder. By training the model to reconstruct the original skeleton sequence, it simultaneously learns to capture the underlying patterns of actions. Extensive experiments on NTU and NW-UCLA datasets demonstrate the state-of-the-art performance as well as the impressive generalizability of our proposed approach. Moreover, our experiments reveal that our approach is capable of generating conditioned skeleton sequences, offering the potential to enhance small datasets or generate samples of under-represented classes in imbalanced datasets. Our code will be published on GitHub.