Digitální knihovnaUPCE
 

Fakulta elektrotechniky a informatiky / Faculty of Electrical Engineering and Informatics

Stálý URI pro tuto komunituhttps://hdl.handle.net/10195/3847

Práce obhájené před rokem 2008 jsou uloženy pouze v kolekci Vysokoškolské kvalifikační práce

Procházet

Search Results

Nyní se zobrazuje 1 - 7 z 7
  • Konferenční objektpeer-reviewedpostprint (accepted version)Omezený přístup
    Medical Catheters Grasping Point Detection with Quality Control
    (SPRINGER INTERNATIONAL PUBLISHING AG, 2021) Štursa, Dominik; Doležel, Petr; Zanon, Bruno B
    The ability to grasp objects is one of the basic functions of modern industrial robots. The emphasis of this paper is placed on the visual perception system, and in particular, on the data processing method leading to grasp point detection. The solution involved the design of a perceptual system in which it was necessary to use a SWIR sensor that can see through plastic bags and thus provide sufficient image information for possible processing by a neural network. The grasping point detection was tested with three convolutional neural network architectures. The method was evaluated by a generalized intersection over union (gIoU). The superior architecture was Attention U-Net, where gIoU reached 0.8522 in the case of the best model.
  • Konferenční objektpeer-reviewedpostprint (accepted version)Otevřený přístup
    Multiple Objects Localization Using Image Segmentation with U-Net
    (IEEE (Institute of Electrical and Electronics Engineers), 2021) Štursa, Dominik; Doležel, Petr; Honc, Daniel
    Precise object localization in an industrial environment is a significant task affecting follow-up processes for a pick and place application. One of the solutions to effectively ensure the success of this task is to use modern methods of machine vision. Machine vision is still a highly evolving topic, in which the use of approaches based on convolutional neural networks is rising. And so in this contribution, an innovative engineering approach based on convolutional neural networks is proposed for an object localization task. The approach is based on an atypical image segmentation, where the individual objects are represented by two colored gradient circles. These circles represent significant parts of the object like its center or ending. Each object type (class) is determined by a specific color. By use of a local maxima finder, all circles in an image are transformed to points. With knowledge of these points the coordinates and rotations are calculated. The proposed approach was tested on a legitimate localization problem with 100% precision, more than 99.52% recall on the positioning task and with an average of 6 minutes angle variance per object.
  • Článekpeer-reviewedpublished versionOtevřený přístup
    Memory Efficient Grasping Point Detection of Nontrivial Objects
    (IEEE (Institute of Electrical and Electronics Engineers), 2021) Doležel, Petr; Štursa, Dominik; Kopecký, Dušan; Jecha, Jiří
    Robotic manipulation with a nontrivial object providing various types of grasping points is of an industrial interest. Here, an efficient method of simultaneous detection of the grasping points is proposed. Specifically, two different 3 degree-of-freedom end effectors are considered for simultaneous grasping. The method utilizes an RGB data-driven perception system based on a specifically designed fully convolutional neural network called attention squeeze parallel U-Net (ASP U-Net). ASP U-Net detects grasping points based on a single RGB image. This image is transformed into a schematic grayscale frame, where the positions and poses of the grasping points are coded into gradient geometric shapes. In order to approve the ASP U-Net architecture, its performance was compared with nine competitive architectures using metrics based on generalized intersection over union and mean absolute error. The results indicate its outstanding accuracy and response time. ASP U-Net is also computationally efficient enough. With a more than acceptable memory size (77 MB), the architecture can be implemented using custom single-board computers. Here, its capabilities were tested and evaluated on the NVIDIA Jetson NANO platform.
  • Konferenční objektpeer-reviewedpostprint (accepted version)Omezený přístup
    Airspace Object Detection Above the Guarded Area Using Segmentation Neural Network
    (Springer Nature Switzerland AG, 2021) Štursa, Dominik; Doležel, Petr; Merta, Jan
    With the increasing number of drones and unmanned aerial vehicles (UAVs), more emphasis is placed on guarding of the airspace around private and also public buildings. In this contribution authors are introducing a complex multi-step approach for aerial objects detection. Introduced process is composed of a few consecutive steps, where objects are cropped from original input with use of cropping pattern provided by task of image segmentation. These objects are then classified and evaluated as a threat or not. However, the emphasis here is placed on the segmentation part only. Neural network topology, adopted from U-Net architecture, was proposed. Case study was made and discussed in an effort to cover a large number of possible states. The results of a proposed convolutional neural network architecture were compared with the U-Net architecture. Applying of the convolutional neural network to the task of airspace object detection lead to sufficiently precise results, thanks to which it is possible to assume the possibility of its use in the proposed multi-step detection system in further work.
  • Konferenční objektpeer-reviewedpostprintOtevřený přístup
    Novel Approach for Person Detection Based on Image Segmentation Neural Network
    (Springer Nature Switzerland AG, 2020) Štursa, Dominik; Baruque Zanon, Bruno; Doležel, Petr
    With the rise of the modern possibilities in computer science and device engineering, as well as with growing population in big cities among the world, a lot of new approaches for person detection have become a very interesting topic. In this paper, two different approaches for person detection are tested and compared. As the first and standard approach, the YOLO architectures, which are very effective for image classification, are adapted to the detection problem. The second and novel approach is based on the encoder-decoder scheme causing the image segmentations, in combination with the locator. The locator part is supposed to find local maxima in segmented image and should return the specific coordinates representing the head centers in the original image. Results clearly report this approach with U-Net used as encoder-decoder scheme with the locator based on local peaks as the more accurately performing detection technique, in comparison to YOLO architectures.
  • Konferenční objektpeer-reviewedpostprintOtevřený přístup
    Counting Livestock with Image Segmentation Neural Network
    (Springer Nature Switzerland AG, 2020) Doležel, Petr; Štursa, Dominik; Honc, Daniel; Merta, Jan; Rozsívalová, Veronika; Beran, Ladislav; Hora, Ivo
    Livestock farming industries, as well as almost any industry, want more and more data about the operation of their business and activities in order to make the right decisions. However, especially when considering very large animal farms, the precise and up-to-date information about the position and numbers of the animals is rather difficult to obtain. In this contribution, a novel engineering approach to livestock positioning and counting, based on image processing, is proposed. The approach is composed of two parts. Namely, a fully convolutional neural network for input image transformation, and a locator for animal positioning. The transformation process is designed in order to transform the original RGB image into a gray-scale image, where animal positions are highlighted as gradient circles. The locator then detects the positions of the circles in order to provide the positions of animals. The presented approach provides a precision rate of 0.9842 and a recall rate of 0.9911 with the testing set, which is, in combination with a rather suitable computational complexity, a good premise for the future implementation under real conditions.
  • Konferenční objektpeer-reviewedpostprintOtevřený přístup
    Convolutional Neural Network for Sound Processing - Study of Deployed Application
    (IEEE (Institute of Electrical and Electronics Engineers), 2019) Doležel, Petr; Štursa, Dominik; Honc, Daniel
    Pest birds are considered as a special kind of vermin, since, in most of countries, their legal position does not enable their direct extermination. Therefore, in order to protect the agricultural areas indirectly from pest birds, the robust and highly selective pest bird sensor is necessary to design. In this contribution, the pest bird detection unit, based on a convolutional neural network, is presented. The convolutional neural network itself is used for the decision making about the pest bird occurrence, while sound recordings are used as input data. The testings, presented at the end of the contribution, proved a very high accuracy of the detection unit, with the results indispensably improved in comparison to previously presented approaches.