System performance analysis and fixed-point architecture of a gradient-based optical flow algorithm

System performance analysis and fixed-point architecture of a gradient-based optical flow algorithm

Hurkmans, T.M.F.

Van Leuken, T.G.R. (mentor)
Van der Meijs, N.P. (mentor)
Van Genderen, A.J. (mentor)
Van der Veen, A.J. (mentor)

Electrical Engineering, Mathematics and Computer Science

Circuits and Systems

Computer Engineering

2009-12-15

Optical flow algorithms present a way for computers to estimate motion from the real world. Applications like cloud motion, surveillance and robot eyesight are examples of this. The focus of existing research is mainly on either fast, but poor solutions, or slow but good solutions. In this thesis an approach to improve performance through reducing computational effort of the Lucas \& Kanade optical flow algorithm is presented. An architecture of this optimised complex algorithm on an FPGA is given. After an introduction into optical flow algorithms, a high-level system design optimisation process of the large complex computing blocks will be described, in which each part of the algorithm is analysed for quality, speed and computational effort. A new smoothing filter, StackBlur, will be introduced for the first time in an optical flow algorithm. All large complex blocks are integrated into a modular high-level pipelined hardware architecture, using manually generated flow graphs and the two-process design methodology, while keeping flexibilty in mind. Complex mathematical operations including a least squares filter and an eigenvalue computation are presented. Functional correctness of the design is verified through functional and post-synthesis simulation using realistic examples. The obtained results show that performance has been improved in terms of quality and computational effort, but that especially the eigenvalue computation block still needs further improvement.

optical flow
fixed point architecture
performance analysis
two process
lucas
kanade

http://resolver.tudelft.nl/uuid:88add684-df53-4f64-968c-2b44eba315d1

2009-12-18

Student theses

master thesis

(c) 2009 Hurkmans, T.M.F.