Publication Details
Low-Level Image Features for Real-Time Object Detection
Zemčík Pavel, prof. Dr. Ing., dr. h. c. (DCGM)
Hradiš Michal, Ing., Ph.D. (DCGM)
Juránek Roman, Ing., Ph.D. (DCGM)
Havel Jiří, Ing., Ph.D. (CM-SFE)
Jošth Radovan, Ing., Ph.D.
Žádník Martin, Ing., Ph.D. (DCSY)
real-time object detection, image features, Local Rank Patterns
The main aim of the chapter is to provide information about the Local Rank Patterns image feature: Its background, mathematical definition, evaluation of its performance and notes on its implementation and use in object detectors. Implementations on the MMX, SSE, FPGA (programmable hardware), GPU (Cg) and CUDA platforms are described and experimentally evaluated. The performance of the image feature is evaluated within the WaldBoost classifier on the task of face detection, and it is compared to the commonly used Haar wavelets, local binary patterns and other low level features. The Local Rank Patterns feature seems suitable for hardware acceleration both directly by programmable or hard-wired hardware, but also by processors supporting different sets of the SIMD instructions. It is shown, that the LRP feature is an important alternative for construction of fast object detectors.
@inbook{BUT55167,
author="Adam {Herout} and Pavel {Zemčík} and Michal {Hradiš} and Roman {Juránek} and Jiří {Havel} and Radovan {Jošth} and Martin {Žádník}",
title="Low-Level Image Features for Real-Time Object Detection",
booktitle="Pattern Recognition, Recent Advances",
year="2010",
publisher="IN-TECH Education and Publishing",
address="Vienna",
pages="111--136",
isbn="978-953-7619-90-9"
}