Thiago Farias,  João M. Teixeira,  Gabriel Almeida,  Pedro Leite,  Veronica Teichrieb,  Judith Kelner. 

This paper describes an enhanced implementationof the Kanade-Lucas-Tomasi feature tracker,enabling real time feature detection and tracking inhigh definition images. The tracker is implementedusing the nvidia cuda as a massively paralleltechnology for computational intensive tasks. cudahas shown to be a promising technology for realtime high definition image processing and computervision algorithms. In this work we explain howthe real time constraint was achieved through theuse of a gpgpu technology. The implementationhas provided speed up improvements of 90x whencompared to the cpu reference implementation. Wewere able to work with high definition resolutions,as 50 fps results were achieved when extracting1,000 features on a 1,024 x 1,024 video stream.

http://www.lbd.dcc.ufmg.br/colecoes/svr/2009/003.pdf

Processando consulta.