Abstract: A Simultaneous Localization and Mapping (SLAM) algorithm based on Gray-EKF was designed by using gray prediction theory and EKF in order to improve the accuracy of SLAM for intelligent agricultural vehicle. A GM(1,1) prediction model of observation based on improved sliding window gray prediction was set up on the foundation of traditional EKF, and then innovations were made out. With the purpose of raising accuracy and capacity of resisting disturbance, a trinocular stereo vision camera was used as the observation sensor and an accuracy calibrating algorithm was brought out simultaneously. Experiments showed that the calibrated trinocular camera was well precision and anti-interference. The errors of 12 measuring results in 16 were less than 1%. The average observation errors were 0.074 and 0.073 m in x and y axis, and position error was 0.140 m with RMSE 0.048 in the park environment of 30 landmarks. The average observation errors were 0.061 and 0.068m in x and y axis, and position error was 0.109 m with RMSE 0.038 in the park environment of 60 landmarks. The average observation errors were 0.079 and 0.077 m in x and y axis, and position error was 0.122 m with RMSE 0.049 in the dry land environment of 60 landmarks. Compared with traditional EKF SLAM algorithm, Gray-EKF SLAM algorithm is more precise.