Detecção de objetos e estimativa de posição 3D através da aplicação de redes neurais convolucionais

Abstract

To successfully apply robotics solutions in uncontrolled environments it is necessary to obtain quality information about the robot’s operating environment. Detecting, recognizing and estimating 3D position of objects are part of this process and are important for tasks such as object manipulation with a robotic arm. Methods based on convolutional neural networks and 3D models of objects were successful in this task, in different scenarios, despite the difficulty of obtaining 3D models in a practical situation. Also, not all methods can be executed in embedded systems, such as those usually present in autonomous service robots. RGB-D sensors have recently gained importance as tools for environment 3D sensing. These sensors associate RGB color images (red, green and blue) with a depth estimate for each pixel of the image, making it possible to estimate the 3D position of objects in the scene with respect to the sensor. This work proposes and evaluates a method that, from two RGB-D shots made from different positions, estimates an object’s 3D center point, without the need of knowing the object’s 3D shape in advance. The experiments show that the average euclidian distance between the estimated and the real 3D center points was 9.4 cm, with a standard deviation of 5.4 cm and in the best estimation this value was 8 mm.