Real Time Computer Vision for Multi Agent Mobile Robot Control

Júnio Eduardo de Morais Aquino¹; Bruno de Paiva Teixeira¹; Luiz Carlos Figueiredo¹

¹ PUC Minas

doi:10.20906/CPS/CILAMCE2017-0981

Resumo

In this work we present our formulation, development and tests of GPU accelerated image analysis techniques for use with a computer vision based aproach to control several mobile robots. The tests were realised using robots designed for the IEEE Very Small Size Robot Soccer category. In this category the robots are required to be have maximum dimensions of 75 milimeters of length, width and height, the images are obtained from a single camera positioned 2 meters above the field. The robots and the ball can be identified by their unique colors. The robots are in a dynamic environment with mobile obstacles, hence, in order to the data gathered from the image analysis to be viable to further processing and decision making it is required a very small processing time for each frame. Using the most recent graphics processors units architectures available for the computer gamming market, we were able to obtain position and velocity estimations from images with great precision and fast processing times, enabling the use of several artificial inteligence techniques and path planning algorithm to control the movements and other actions of the robots. The position of the robots and obstacles was obtained using color thresholding and the velocity estimated using an Lucas-Kanade based optical flow algorithm.

Palavras-chave: Computer Vision; Real Time Image Analysis; GPU accelerated processing; Optical Flow; Object Detection; Mobile Robots