In order to achieve small agricultural automated guided vehicle (AGV) which could navigate autonomously between corn rows, a method was proposed which could quickly and accurately extract the centerlines of middlelate corn rows as the innovative methods of navigation baseline of agricultural robot in field environment. The algorithm was improved by the traditional 2G-R-B algorithm so that it could obtain the characteristics of green plants in real-time and also improve the robustness of image pre-processing. According to the vertical projection of corn crop’s line, points of profile features of corn roots were generated. Using the detection algorithm of peak points, the backup location points of corn roots were obtained and then the location points of corn roots were got after the second judgment and detection. The least square method was used for fitting the location points of corn roots and two lines of crop rows were generated. The actual navigation baseline was calculated based on the formula angle bisector after the line slopes of two crop rows were generated respectively. In addition, the camera calibration process was simplified so that the image pixel coordinates could be converted into world coordinates quickly. The extracted angle and lateral deviation of navigation baselines were used as input parameters of navigation to control agriculture AGV. The experimental results showed that the method had strong robustness which could adapt to different environments and the accuracy of navigation baselines by detecting was more than 90%. The average processing time of a 700 pixels × 350 pixels color image was less than 185 ms which had a better real-time. The results provided a reliable reference method for autonomous navigation of the agricultural AGV in middle-late corn field.