Page - 123 - in Proceedings - OAGM & ARW Joint Workshop 2016 on "Computer Vision and Robotics“

the whole procedure, the error of the row distance estimation ed is directly determined based on manually measured ground truth data. The particle filter is initialized withN=1000 randomly gen- erated hypotheses that represent parameter configurations withα= [−pi 2 , pi 2 ] , p=[−0.75m,0.75m], andd=[0.2m,1.5m]. Figure6: FRANC during in-field trials. x y dd p 321 Figure7: Crop rowsandparameter windows. The parameter windows are defined with pw=[0.2m,−0.2m], αw=[+0.2rad,−0.2rad], and the manually measured ground truth data for the row distancedGT=0.45m. The experiments show that the particle filter based crop row detection ends in average after five cycles in correct estimations for all three parameters (cf. Fig. 8). The steps within the row offset can be ascribed to the normalization algorithm that searches for the closest line of the pattern to the origin of the coordinate system that was slightly shifted to the right side during the recordings. Hence, the orientation and the offset of the rowpattern is eitherdescribed with line 2©or 3© (cf. Fig. 7). 0 50 100 150 −0.1 0 0.1 Offset Plot t [s] 0 50 100 150 −0.2 0 0.2 Orientation t [s] 0 50 100 150 −0.1 0 0.1 RowSpace Error t [s] Figure 8: Results of the row detection algorithm with data recorded during in field trials. Offset and orientation has to be within the windows as stated in the text. Average error of the row distance parameter refered to theground truth. 5. Conclusion In thisarticlewepresentedadesignconcept foramodularagricultural robotand its realisation in the FRANCprototype including resultsonpreliminaryfield trials. TestingFRANCinthefieldproveditsmaneuverabilityonroughterrain. Therecordedin-fielddatafor theevaluationof the rowguidancealgorithmrevealed that theparticle-filter-basedcroprowdetection ends inaverageafter five cycles incorrect estimations. We believe that the conceptual design, its prototypical realization, and the preliminary field trial results presented in this article constitute valuable knowledge for fellow researchers in the field of agricultural robotics and serve as a stepping stone towards developing robotic modules for more flexibleagricultural automation. 123