A Comparative Assessment of Parcel Box Detection Algorithms for Industrial Applications

Industrial logistics may benefit from object perception to perform flexible and efficient management of goods. This paper illustrates and experimentally compares two approaches to parcel box detection in depth images for an industrial depalletization task. The model-based method detects clusters in the input point cloud according to curvature and other geometric features, and aggregates the candidate objects. The learning-based method relies on the state-of-the-art Mask R-CNN, which has been re-trained on an acquired dataset with missing measurements. The target object poses are evaluated through standard geometric registration. The experiments on acquired datasets show the feasibility of the two approaches.