More organic through agricultural automation
A hundred years ago, Germany was still an agrarian state: 38 percent of the labour force worked in agriculture. A farmer produced food in order to feed an average of four people. Nowadays things are completely different: two percent of the labour force work in agriculture, feeding an average of 131 people each. The numbers from Germany from 2010 reflect the general picture of the agricultural industry in industrialised nations. This has essentially been made possible by the technological progress made in the agricultural sector.
Changing consumer behaviour also means that the end of automation in agriculture remains some way off. Organic is on everyone's lips and this is confirmed by the annual growth rates for organic food of between 7 and 10% (in Germany). The use of herbicides for the chemical treatment of weeds is not part of the green picture, even if the quantity were to be reduced to a minimum through considered use. Manual weeding is also not an option for reasons of cost and time. This was a view shared by the French company Naïo Technologies, which is based near Toulouse, so it started work on developing a weeding robot.
Two calibrated cameras connected in stereo ensure precise positioning. Calibrated means that both the cameras are calibrated to each other (intrinsic) and that the stereo connection is calibrated to the "real" world by a translation vector and a rotation matrix (extrinsic). If the system now compares corresponding points in two successive stereo recordings, it can use triangulation to calculate the change in the Z axis in order to obtain the movement information. To keep the calculation of data to a minimum, the image is divided into rectangles and reduced to a small number of properties in the image. To ensure that the system remains robust in spite of this, "information reduction", outliers are removed from the measurements using the RANSAC algorithm.
To extract the navigable path, Naïo additionally uses a stereo correspondence algorithm, which computes a disparity map of the stereo camera pair. In combination with stereo-matching, the robot can also detect obstacles. Both systems work at different frequencies, i.e. the odometry system with 15 frames per second, the stereo correspondence algorithm with 5 frames per second. Using both results, Naïo reconstructs a 3D map which the robot can keep in the memory.
When selecting the hardware components, Naïo decided on power-saving embedded components linked to a compact Linux system. So far, so good, but the first series revealed a large number of bugs in the Linux driver for the selected camera, with the result that simultaneous triggering of both sensors seemed almost impossible. As a result, Naïo continued its search for a suitable camera, finding it in the MATRIX VISION USB 2.0 board camera mvBlueFOX-MLC200w. The camera is equipped with an Aptina Global-Shutter CMOS sensor, the high sensitivity of which makes it ideal for outdoor use in variable light conditions.
The switch to the MATRIX VISION camera paid off from the very start. The direct contact with development and support and the quick response to identified bugs allowed the camera to be integrated more quickly than anticipated. The synchronous triggering of several mvBlueFOX-MLC board cameras is one of the basic functions of the camera, which also impressed the developers at Naïo, who were able to configure the cameras quickly thanks to the good documentation provided. According to the developers, another very useful feature is the option of uploading the calibration settings for the entire system to the respective camera's non-volatile memory. This makes the system more robust and provides a dual backup for the calibration.
Jean Inderchit, Robotics Engineer at Naïo Technologies
"The MATRIX VISION driver operates reliably on all platforms. This means that Naïo remains flexible in the event of future platform changes and can continue to build upon the MATRIX VISION cameras.