Page - 122 - in Proceedings - OAGM & ARW Joint Workshop 2016 on "Computer Vision and Robotics“

(a) (b) (c) (d) Figure5: Comparison of different segmentation methods. (a) RGB image, (b) 2G-R-B segmentation [15] , (c)NIRsegmentation, (d)NIRDsegmentation. (a), (b), (c), and (d) showthesamesceneunder different fieldofviews. of the closest line and points to the origin of the coordinate system. The third parameter is the scalar dwhich describes the distance between the lines of the repetitive pattern. The filter samples a 3D parameter space with N hypotheses. Each hypothesis is weighted based on the segmented image. In opposite to other methods the approached crop row detection does not need any prior information on the row structure. Moreover, the particle-filter-inherent properties in combination with the selected geometric row model enable a tracking of the crop rows and improve the results even and especially ifnatural rowirregularitiesoccur. Finally, thenegotiable track isextractedoutof the rowinformation and is further filtered and processed for the steering information. To achieve the modularity of the whole system, the rowguidance iswrapped in the robotoperatingsystem(ROS)andcanbe replaced by another guidance system if necessary. In our recent work we have investigated in [6] how the fusionofodometryand rowguidance informationcan improve thedetection results. 4. TestsandResults As proof of concept we built with the developed subsystems the robotic platform FRANC (cf. Fig. 6). It consists of a frame that carries the electronic and sensor system and is powered with four independent steerable wheels. The algorithms, controller, and the security concept including the remote control with the emergency stop function were implemented to form a whole system with minimaleffort. As stated by [13] the integration task can be a significant effort on its own. The modular concept reduced the integration of the single modules into an overall system to a few mechanical engineering steps as the preparation of the frame including the mounting points and an one-time parametrization of theelectricalsystemandthecontrolalgorithms. Theparameterizableandadaptablealgorithmsand interface design simplifies the integration of the subsystems into a working solution and overcomes several integrationproblems thathave tobe faced in traditionally designedsystems. FRANC was successfully tested in rough terrain and recorded in-field data for the evaluation of the row guidance algorithm that is used by the autonomy software. The tests proved the feasibility, maneuverability, and rigidityofour modular concept for real-life applications. The crop row detection algorithm and row guidance software is tested with data recorded during in-field tests of the robot. The robot was maneuvered within row organized fields, parallel to the rows. With this information, parameter windows for p andα can be defined to evaluate the crop row detection algorithm. Correct row structure estimations have to end in a parameter configura- tion that describe rows within the given windows. Since the row distance has to be constant during 122