HR4RS stamp
Irene Vega
The developed system allows automatic recognition of human activities and hand gestures, such as catching, in images recorded with depth cameras. The results obtainedpublished in the scientific journal Pattern recognition, are one more step towards making machines understand human behavior better. "The recordings, made with Microsoft Kinect-type cameras, provide us with 3D information on the position of the joints of our skeleton," explains Juan Carlos Nuñez, one of the authors of the study.
This new technology can find multiple applications in the detection of suspicious activities in automatic video surveillance systems, in the labeling of the activity carried out by an athlete in a sporting event or in the recognition of the action that a person is carrying out in a video game. .
This work is part of a line of research in which scientists from the groups GAVAB (Group of Algorithms applied to Artificial Vision and Biometrics) and HOOD (High Performance Computing and Optimization) of the URJC. The advances achieved in some cases exceed existing methods, reaching success rates approaching 100%. "We have carried out experiments on the main publicly available test bases, such as MSR Action3D, or UTKinect-Action3D, obtaining a success rate of 95,7 and 99%, respectively," says Juan Carlos Nuñez.
Mathematical models to emulate neural networks
The tools used for the design of this new technology are based on the use of artificial neural networks. These systems consist of mathematical models that simulate some behaviors of biological neural networks, such as their learning capacity.
According to Juan Carlos Nuñez, “with recent advances in artificial neural network learning techniques, it has been possible to create deep networks, which consist of many layers of stacked neurons. In addition, due to their great learning capacity, they are being applied very successfully to a multitude of problems that were traditionally considered difficult, such as recognizing objects in photographs or recognizing speech”.
The work carried out by the URJC researchers also presents a new data augmentation strategy, which has allowed them to extend the limited number of examples available for the network learning process. "This strategy has been fundamental, because deep neural networks require a large number of examples to learn", explain the authors of the article.
This study is part of the doctoral thesis of Juan Carlos Nuñez, supervised by Raúl Cabido, member of the CAPO group, and José Vélez, member of the GAVAB group. Professors Antonio Sanz and Juan José Pantrigo have also participated. The study is part of several projects, within the calls of the National R&D Plan, and has funding from the Banco Santander and URJC program for excellence in research for the group Computer Vision and Image Processing (CVIP).