Many real-time vision systems working in complex and uncertain environments need a robust algorithm that can successfully perform the line-segment extraction. Recently, Lee & Kweon [1] proposed a method which can find out lines without any assumptions and constraints in variety of real scenes. However, the computational burden of this algorithm makes it difficult to be applied to real-time applications. This thesis develops some techniques to reduce computation time and achieves a real-time implementation by using a general purpose DSP.
Since floating-point operations take much computation time compared with other operations in a usual DSP, integer operations should be used as much as possible. In addition, LUT (Look-Up Table) based approaches often drastically reduce computational complexity. To the end, this thesis realizes an 5-times faster version of Lee's algorithm using integer operations and LUT approaches. Also, in order to achieve a real-time implementation of the proposed algorithm, a TMS320C80 based DSP board has been used. In a series of experiments for outdoor traffic scenes, the proposed system has processed 5-10 frames per second depending upon the complexities of scenes. Substantial algorithm is coded in assembly language. The proposed method can provide a fast and reliable line extraction capability.
