Reliable and accurate visual detection of crop rows is prerequisite for implementing successful autonomous navigation for plant protection robots. A visual navigation path detection approach based on random sample consensus (RANSAC) algorithm was proposed. Firstly, the excess green (ExG) method and the maximum variance between classes were used to figure out gross target regions. Secondly, morphological operations and dynamic area threshold filtering strategy were employed to filter out the interferences. As outlier points significantly influenced the estimation accuracy, RANSAC algorithm was proposed to purify the inlier point sets. Finally, crop rows line features were modelled by least mean square techniques, which offered a degree of robustness in constructing global co-linear features in contrast to Hough transformation. To sufficiently verify the effectiveness of the idea, wheat, peanut, corn and film covered potato seedling images were used for evaluation. As revealed by experimental results that the proposed method outperformed Hough transformation in the crop rows center line extraction, and RANSAC algorithm rendered the method more robust with respect to noise and outliers, which allowed the successful detection rate of the work to be improved by 18.8 percentage points and arrived at 93.8%. The overall framework made sense to reliable visual navigation for plant protection robots.