FASTDLO: Towards Real-Time Perception of Deformable Linear Objects

In this paper is presented an approach for fast and accurate segmentation of Deformable Linear Objects (DLOs) named FASTDLO. The perception is obtained from the combination of a deep convolutional neural network for the background segmentation and a pipeline for the dlo identification. The pipeline is based on skeletonization algorithm to highlights the structure of the DLO and a similarity-based network to solve the intersection. FASTDLO is trained only on synthetically generated data, leaving real-data only for evaluation purpose. FASTDLO is experimentally compared against DLO-specific approach achieving better overall performances and a processing rate higher than 20 FPS.