libHOG: Energy-Efficient Histogram of Oriented Gradient Computation

Histogram of Oriented Gradients (HOG) features are the underlying representation in automotive computer vision applications such as collision avoidance and lane keeping. In these applications, the authors have observed that HOG feature computation is often a slow and energy-intensive component of the overall pipeline. In this paper, the authors focus on reducing both the time taken and the energy used for computing Felzenszwalb HOG features. The authors achieve the results though a combination of reduced precision, SIMD parallelism, algorithmic changes, and outer-loop parallelism. In particular, the authors address a bottleneck in histogram accumulation by phrasing the problem as a gather instead of the (traditional) scatter. Additionally, the authors explore the tradeoffs of using L1 instead of L2 norms to compute gradients, which enables smaller operands and more SIMD parallelism. Overall, the authors are able to compute multiresolution HOG pyramids at 70fps for 640×480 images on a multicore CPU. This is a 3.6x speedup over the best known HOG implementation and a 29× speedup over the popular voc-release5 HOG code. This is also a 3.6× - 22× reduction in energy per frame compared to previous HOG implementations. The authors' open-source implementation is available for download.


  • English

Media Info

  • Media Type: Web
  • Features: References;
  • Pagination: pp 1248-1254
  • Monograph Title: 18th International IEEE Conference on Intelligent Transportation Systems (ITSC 2015)

Subject/Index Terms

Filing Info

  • Accession Number: 01602921
  • Record Type: Publication
  • ISBN: 9781467365956
  • Files: TRIS
  • Created Date: May 2 2016 3:20PM