Background: Vision sensors have been widely used in mobile robot navigation nowadays. Research on autonomous navigation of mobile robots for tree rows following in groves and orchards is rare relative to crop rows. In this paper, a vision navigation subsystem for an autonomous vehicle operating in citrus groves was developed. Methods: Tree rows were first extracted according to colour difference using a HSV colour transformation. A specific algorithm was applied to remove the noise caused by the weeds on the driving path and the horizontal scanning method was used to detect the boundaries. The boundary lines were detected by the least squares method combined with RANdom Sample Consensus (RANSAC) mechanisms. A guidance directrix was then generated based on these boundary lines. Results: The average processing time for one image was about 70ms. All offset errors were compared with the reference trajectory obtained by manually selecting from the image. The average MSE offset error from 261 off-line orange grove images was 4.53 pixels with an average orientation offset error of 2.14 degrees. The average MSE offset error from 631 images which were captured from real-time vehicle experiments was 3.1 pixels with an average orientation offset error of 1.95 degrees. Discussions: The off-line image sequences and real-time processing experiments showed that the algorithm was able to overcome weeds noise problem on the ground, generate the guidance directrix successfully and met real-time processing requirements. Conclusion: In this paper, a machine vision algorithm to obtain a guidance directrix for an automatic vehicle guidance system is presented. The algorithm could effectively overcome noise problems derived from weeds on the ground. Further research would be interested in improving the applicability of the algorithm under a range of weather and lighting conditions and other kinds of grove and orchard.