PosePP: 6D Pose Estimation of Poultry Pieces in a Cluttered Bin

Pose estimation of deformable natural food products is challenging due to the large variation in shape and sizes among objects. This is especially true for bin picking, which additionally involves clutter and occlusion. This paper proposes a novel modular algorithm for 6D pose estimation of chicken legs in a cluttered bin picking scenario using an RGB and depth (RGB-D) image. We use a Mask R-CNN implementation for the instance segmentation of objects, that produces region-of-interest (ROI) proposals and simultaneously predicts intermediate mask representations of the objects. The proposed 6D pose estimation algorithm uses principal component analysis (PCA) utilizing the output mask from Mask R-CNN and the depth image of the objects. The algorithm can detect and classify fully visible pieces with a [email protected] of 0.93. The average error of position of the chicken legs was measured at 3 mm and the rotation error of chicken legs was estimated with an error of 1. 33 ± 1.31 degrees. Overall, the algorithm could correctly estimate principal axes for 89% of the cases. Experimentally, we show that the proposed algorithm can detect and estimate the pose of the chicken legs.