A real-time human action description processor with spatio-temporal HMAX model = 시공간 기반의 계층적 인식기를 이용한 실시간 인간 행동 기술 프로세서

Cited 0 time in webofscience Cited 0 time in scopus
  • Hit : 414
  • Download : 0
In these days, human action recognition is applied to various applications such as gesture recognition embedded in some mobile devices and video retrieval which is a searching technique for users to find a specific part containing acting human. For these applications, human action recognition is asked for having both real-time performance for good user interactivity and high action recognition accuracy. A bio-inspired hierarchically description model named HMAX is known as one of the most accurate human action description algorithms, whereas its too much slow processing speed is a problem to be applied for many applications mentioned above. Generally speaking, 30 FPS(Frames Per Second) processing speed is required for comfortable user interaction. However, it is figured out that HMAX has about 1 FPS processing speed from software simulation with general purpose CPU. Its slow processing speed comes from the template matching stage called `second simple stage` because of two main factors; (1) number of matched templates, and (2) SAD processing time per a template. The total processing time of this stage is mainly proportional to them, and they should not be ignored. This is the reason why template matching stage is the bottleneck of overall HMAX processing. For reducing critical burden of the template matching stage, two algorithm modification methods are introduced. One is spatio-temporal ROI(Region-Of-Interest) processing which reduces unit SAD processing time per a template. This method reduc-es SAD processing time per a template by 69.33%. Another is class pre-selection which reduces total number of matching templates. This method reduces total number of template by 93.98%. Two modification methods accompanies with additional computation cost. For accurate spatio-temporal ROI generation, fast conversion operation between 2-dimensional position information and internal address of SRAM is needed. In addition, multi-dimensional sorting operation should be suppo...
Yoo, Hoi-Junresearcher유회준
한국과학기술원 : 전기및전자공학과,
Issue Date
569238/325007  / 020123274

학위논문(석사) - 한국과학기술원 : 전기및전자공학과, 2014.2, [ iii, 32 p. ]


Human action description; 실시간 프로세서; 시공간 기반; 계층적 인식기; 인간 행동 기술; Real-Time Processor; HMAX; Spatio-Temporality

Appears in Collection
Files in This Item
There are no files associated with this item.


  • mendeley


rss_1.0 rss_2.0 atom_1.0