Depth Based Fruit Detection from Viewer-Based Pose

Seeking to accurately detect and localize fruit of any color in 3D space for selective agrobotical operations, we exploit data given by Time-of-Flight or RGB-D cameras and propose a novel shape-based fruit detector using a fruit pose reference frame relative to the viewer. Surface normals, which are shaped-based local features, are accumulated into bins of different shapes along the reference frame's axes. The used normals are represented using two angles from a viewer-based reference frame, to achieve a representation that is suitable for fruit types that are almost symmetric around an axis (e.g., bell-peppers), without having an effect on fruit types with no axis-symmetry. Results are shown on a particularly challenging pepper dataset.