Page - 121 - in Proceedings - OAGM & ARW Joint Workshop 2016 on "Computer Vision and Robotics“

real time control system electricplatformsystem PLC motor drivern motordriver1 I/O encoder 1..n planing moule PC ROS stereo camera NIR camera sensors battery actuators motors remote control Figure4: Electrical platformsystemand theadjacent systems. 3.2. ElectronicsandControlSystem The vehicle electronics is the bridge between the robot kinematic, including the motors, and the autonomyandrowguidancesoftware. Figure4showsanoverviewof thesystemparts. Thenecessary sensors system is closely connected to the implemented row guidance system. Based on the review of the prior work [13, 11] we consider that vision systems provide the information for an adaptable navigationand infield taskexecution. Hence,weapproachavisionsystemthatobserves lightwithin different ranges of the electromagnetic spectra and is mounted on the robot front. The sensor system consists of two stereo cameras and a NIR camera. A NIR pass filter and the sensitivity of the built in chip formincombinationabandpass filter that enablesadetectionof light from850nmto1000nm. 3.3. RowGuidanceandAutonomySoftware The row guidance system consists of a segmentation step, followed by a detection of the rows and a parameter extraction. The images are segmented based on NIR and depth data that are provided by the camera system [7]. The extraction of the height information is realised with an online plane calibration that allowsdetermining thecamerapose relative to theestimatedgroundplane. Several machine vision based row guidance approaches [1, 8, 12] consider pure RGB or NIR infor- mation for the segmentation of the plants and soil, while 3D information is omitted and the other way round [9, 14]. Pure RGB-data-based segmentations often fail to segment crops from the soil if they stopped already the production of chlorophyll and lose their green color, while NIR light is still reflected by the cell structure of the leaf (cf. Fig. 5 (b) and (c)). Otherwise, a pure height-based segmentation fails e.g. in early growing stages of the plants, the spectral information can be used as soon as small plants are visible. We approach in [7] a segmentation that fuses both, NIR and depth information together and utilizes the advantages of the one method to compensate the shortcomings of theother. Theheight information improves the results especially forfieldswhereplants are sowed on dams and allows to filter out small plants and weeds that would add noise to the segmented im- age (cf. Fig. 5 (d)). Further, the available 3D information enables a projection of the segmentation result to the online estimated ground plane and enables a height-bias-free crop row detection. The rowguidancesystemdetects the rowsbasedonageometric rowmodelandaparticle-filter-basedrow parameter estimation as approached in [7]. The row model describes with three parameters a parallel pattern of lines in the 2D space. The first two parametersα andp represent the 2D normal vectorp 121