A novel hardware accelerator for embedded object detection applications

David Watson, Gordon Morison*, Ali Ahmadinia, Thomas Buggy

*Corresponding author for this work

    Research output: Contribution to journalArticle

    Abstract

    Object detection applications often require the algorithms to execute on embedded processing platforms such as Multiprocessor SoCs (MPSoCs). One way these algorithms can search input images for objects-of-interest is by consulting a detection library that contains a list of features describing the objects. The processing of large volumes of image data and consultation with a library can decrease the performance of processing platforms, as contention for cacheable resources leads to varied data locality and reuse: software-based techniques have been investigated in the literature with varied success. This paper addresses this issue head-on through a novel hardware accelerator designed to overcome the disadvantages of shared resources contention, whilst optimising on-chip memory consumption. Detection libraries are compressed and stored on-chip within the accelerator that decompresses the data and writes it to dedicated dual-port memories ensuring optimal library data locality and reuse for all processors. By allowing the accelerator to manipulate library data, application performance can be improved by reducing the computation carried out by processors. Our evaluation revealed that by eliminating contention within caches, application performance was drastically improved without over-consuming on-chip resources or power.
    Original languageEnglish
    Pages (from-to)551-562
    Number of pages12
    JournalIEEE Transactions on Emerging Topics in Computing
    Volume5
    Issue number4
    Early online date25 Jan 2016
    DOIs
    Publication statusPublished - Dec 2017

    Keywords

    • algorithms
    • object detection
    • detection library

    Fingerprint Dive into the research topics of 'A novel hardware accelerator for embedded object detection applications'. Together they form a unique fingerprint.

  • Profiles

    Cite this