When applying Hough transform to straight line detection, a straight line's mapping model can easily be interfered by the other lines, short segment noise or its own un-ideality in the parameter space, which brings invalid votings leading to problems such as fault detection, missed detection and inaccurate endpoint location. A novel method was proposed which introduced Least Square Method (LSM) performed in ρ-θ domain to Random Hough Transform (RHT) algorithm for detecting straight lines. The validity of sample was verified by pixel-length ratio before each voting in order to get rid of pseudo lines, which was followed by linear fitting based on least square method in parameter space for parameter correction. By setting an accumulation threshold, straight line candidates were picked out one by one via detecting peak point. Endpoints of straight line segments were located by setting a gap-and-scale threshold. The method is proved to have higher detecting precision than conventional Hough transform.