Through a computational algorithm, a team of researchers have developed a neural network that allows a small robot to detect different patterns, such as images, fingerprints, handwriting, faces, bodies, voice frequencies and DNA sequences. Nancy Guadalupe Arana Daniel, researcher at the University Center of Exact and Engineering Sciences (CUCEI) at the University of Guadalajara (UDG) in Mexico, focused on the recognition of human silhouettes in disaster situations.
Arana Daniel devised a system in which a robot, equipped with a flashlight and a stereoscopic camera, obtains images of the environment and, after a series of mathematical operations, distinguishes between people and debris. During the imaging process, which has a similar appearance to that of animated films, HD cameras are used to scan the environment. Then, the image is cleaned and the patterns of interest — in this case human silhouettes from the rubble — are segmented.
A friction crawler-based drive system (such as the one in war tanks), ideal for all types of terrain, is necessary to break into irregular ground. The robot also has motion sensors, cameras, a laser and an infrared system, allowing it to rebuild the environment and, thereby, find paths or create 2-D maps.
Due to its complexity, this interdisciplinary project required the support of Alma Yolanda Alanis García, Carlos Alberto López Franco and Gehová Lopez all from CUCEI, who handled both the visual and motion control of the robot. Initially, the whole system is integrated in the robot. However, when the model is too fragile to carry a computer, the algorithm runs on a separate laptop, and the robot is controlled wirelessly. In this way, the human recognition images obtained by the cameras of the robot are transmitted to the computer, said Arana Daniel.
Once the robot obtains silhouettes, it uses the descriptor system that obtains the visual characteristics (3-D points) to segment (wrapping objects in circles), and then it builds the human external silhouettes. These silhouettes will serve as descriptors to train a neural network called CSVM, developed by Arana Daniel, to recognize patterns
After that, the captured images are transformed into numerical values representing shape, color and density. When merged, these figures give rise to a new image, which passes through a filter to detect whether it is a human silhouette or not.
“Pattern recognition allows the descriptors to automatically distinguish objects containing information about the features that represent a human figure; this involves developing algorithms to solve problems of the descriptor and assign features to an object,” explained the pattern recognition specialist at the UDG.
Finally, the purpose is to continue working with the robot and to train it to automatically classify human shapes from previous experience. The idea is to mimic the learning process of intelligent beings, allowing it to automatically relate elements.