The following paper presents the development of an algorithm, in charge of detecting, classifying and grabbing occluded objects, using artificial intelligence techniques, machine vision for the recognition of the environment, an anthropomorphic manipulator for the manipulation of the elements. 5 types of tools were used for their detection and classification, where the user selects one of them, so that the program searches for it in the work environment and delivers it in a specific area, overcoming difficulties such as occlusions of up to 70%. These tools were classified using two CNN (convolutional neural network) type networks, a fast R-CNN (fast region-based CNN) for the detection and classification of occlusions, and a DAG-CNN (directed acyclic graph-CNN) for the classification tools. Furthermore, a Haar classifier was trained in order to compare its ability to recognize occlusions with respect to the fast R-CNN. Fast R-CNN and DAG-CNN achieved 70.9% and 96.2% accuracy, respectively, Haar classifiers with about 50% accuracy, and an accuracy of grip and delivery of occluded objects of 90% in the application, was achieved.

The following paper presents an algorithm for sorting up to 5 different tools based on deep learning and specifically in a convolutional neural network, according to the top in pattern recognition found in state of the art and compared by a Haar classifier in object recognition tasks. A Faster R-CNN is used to detect and classify tools located randomly on a table and a Haar classifier to detect other tools delivered by the user. The Faster R-CNN allows recognizing the existing tools on the table and where they are located in the physical space. The Haar classifier detects and tracks, in real-time, a tool delivered by the user's hand to sort it on the table, together with the other elements. Both the training of the convolutional network and the design of the Haar classifier are exposed. The algorithm detects and classifies the tools found on a table, then orders them side by side, and finally waits for the user to deliver some of the five missing tools on the table, take it from his hand, and locate it at the end of the row of objects. A Faster R-CNN was used with an accuracy of 70.8% and a Haar classifier with a 96% recognition, managing to order the five tools in a physical environment. The average time in comparison demonstrates that the Haar classifier presents a lower computational cost.