In the present work a system for hand detection in depth images is proposed. To perform this task, the scene is segmented into 5 different classes: head, arms, body, hands and background, using the pixel-wise classification of random decision forests. Once the scene is segmented, the connected components algorithm is applied in order to group sets of pixels of the same class into regions. From these regions, a list of hand candidates is generated by validating the obtained components. Using Dijkstra’s algorithm, the points at a geodesic distance of up to 50 cm from the center of the hand candidate are found, along the found route, a histogram of classes is generated and used as a descriptor for the final classification, which is performed with support vector machine. With this proposal, a recognition rate of 83,05 % is reached over a data base of 80000 synthetic images

Proyecto de Graduación (Maestría en Ingeniería Electrónica) Instituto Tecnológico de Costa Rica, Escuela de Ingeniería en Electrónica, 2015.