No abstract

No abstract

This paper presents an object recognition module development. This module uses a local feature approach to identify keypoints in free form objects and an unsupervised artificial neural network (ANN) to associate the nearest ones and get clusters of each object learned. The module uses A-KAZE feature descriptor and Growing Cell Structure (GCS) ANN. The module is validated using an own data base, with twenty real objects and twenty different images each one. Here is presented a variety of experiments using from five to fifteen trainning images per object and the rest of them for evaluation. This method gets good results with 100% of discrimination between objects and up to 80% of correct classification.

<p>Despite autonomous navigation is one of the most proliferate applications of three-dimensional (3D) point clouds and imagery both techniques can potentially have many other applications. This work explores urban digitization tools applied to 3D geometry to perform urban tasks. We focus exclusively on compiling scientific research that merges mobile laser scanning (MLS) and imagery from vision systems. The major contribution of this review is to show the evolution of MLS combined with imagery in urban applications. We review systems used by public and private organizations to handle urban tasks such as historic preservation, roadside assistance, road infrastructure inventory, and public space study. The work pinpoints the potential and accuracy of data acquisition systems to handled both 3D point clouds and imagery data. We highlight potential future work regarding the detection of urban environment elements and to solve urban problems. This article concludes by discussing the major constraints and struggles of current systems that use MLS combined with imagery to perform urban tasks and to solve urban tasks.</p>

Resumen. En este trabajo se presenta el desarrollo de un algoritmo para la construcción de mapas bidimensionales mediante odometría inercial y elementos visuales. Se hace uso de un módulo de reconocimiento de objetos basado en características locales y redes neuronales artificiales no supervisadas. El módulo se utiliza para aprender los elementos no dinámicos en una habitación y asociarles una posición. El mapa queda representado como una red neuronal a la cual cada neurona le corresponde una posición real. Los experimentos se realizaron mediante simulación en Webots y con un robot NAO virtual. Una vez construido el mapa sólo basta con capturar un par de imágenes del entorno para estimar la ubicación del robot. Los resultados demuestran que los mapas bidimensionales alcanzan una precisión de hasta ±(0,06, 0,1) m.Palabras clave: elementos visuales, mapas bidimensionales, odometría inercial, robot humanoide NAO, A-KAZE, GCS.Abstract. This paper presents a map construction algorithm development by inertial odometry and visual features. It uses an object recognition module based on local features and unsupervised artificial neural networks to learn no dynamic elements in a room and assign them a position. The map represent a neural network where each neuron is a real position in the room. The experiments were made by simulation in Webots environment using the virtual humanoid robot NAO. Once the map is built, it is enough to capture a couple of images from the environment to estimate the location of the robot. The results show a good precision in localization with the two dimensional maps through ±(0,06, 0,1) m.