A good GPS sensor (RTK-GPS) can provide as low as 1 cm error of horizontal position [4], but only under good conditions. Error sources are the numbers of satellites visible, signal multi-path and dropouts. This position is also global, and not directly related to the crops measured in the images from the camera. A camera sensor and an algorithm comparing consecutive frames, also known as visual odometry, can provide this non-contact measurement. There are different methods to determine the position. Work on visual odometry has been done with both mono and stereo cameras and with cameras mounted at different angle to the ground [5]-[9]. The feasibility of a visual odometer system has been shown in [5], where high slip terrain on Mars surface makes it hard to navigate using wheel encoders. Since there is no system for global positioning such as the GPS system on Mars, the visual odometry has become a critical vehicle safety system. A stereo camera setup is used and Harris corners are detected as features, which are matched and
tracked using RANSAC with least square estimator for outlier rejection. Other work using forward looking stereo cameras for visual odometry are presented in [6]. Reported error was less than 3% without filtering and less than 1% in combination with other sensors (IMU). In [7] different algorithms are evaluated for visual odometry using a single camera pointing perpendicular to the ground. This method assumed planar ground, which makes this solution sensitive to changes in the height of the camera above the ground. Reported error was 0.3% on planar ground and up to 7.7% on uneven grass due to calibration error.