Hier finden Sie wissenschaftliche Publikationen aus den Fraunhofer-Instituten.

A configurable framework for hough-transform-based embedded object recognition systems

: Sarcher, Julian; Scheglmann, Christian; Zoellner, Alexander; Dolereit, Tim; Schäferling, Michael; Vahl, Matthias; Kiefer, Gundolf


Institute of Electrical and Electronics Engineers -IEEE-; IEEE Circuits and Systems Society; IEEE Computer Society:
IEEE 29th International Conference on Application-Specific Systems, Architectures and Processors, ASAP 2018 : ASAP conference 2018; Milan, Italy, 10-12 July 2018
Piscataway, NJ: IEEE, 2018
ISBN: 978-1-5386-7479-6
ISBN: 978-1-5386-7480-2
International Conference on Application-Specific Systems, Architectures and Processors (ASAP) <29, 2018, Milan>
Conference Paper
Fraunhofer IGD ()
heterogeneous computing; object recognition; computer vision; Guiding Theme: Digitized Work; Research Area: Computer vision (CV)

Real-time object recognition on low-power embedded devices is a widely requested task, needed in manifold applications. However, it is still a demanding challenge to achieve desired performance goals. For example, for advanced driver assistance systems (ADAS) or autonomously driven cars, object recognition and lane detection are indispensable tasks. Another field of application is the continuous retrieval of the construction progress on-site for validation of the construction site status, by detecting installed components using a given CAD model. This paper presents a framework for highly customizable object detection systems implemented on a single heterogeneous computing chip leveraging FPGA logic and standard processors. The FPGA logic is used to implement a custom variation of the Hough Transform and further image processing tasks efficiently. The dedicated logic is supplemented with a software stack, which consists of a Linux operating system, including hardware access drivers, as well as high-level libraries like OpenCV and Robot Operating System (ROS) - all running on the same device. The capabilities of the system are demonstrated for three application scenarios, namely race track recognition, lane recognition and object detection tasks performed within a construction assistance system.