Criado em 2005 com aprovação do primeiro projeto de pesquisa do grupo apoiado pelo CNPq na área de visualização de informação. O grupo tem como características principais trabalhar com tecnologias inovadoras, que possibilitem no avanço e divulgação do conhecimento cietífico e desenvolver produtos aplicados em benefício da sociedade. As Linhas de Pesquisa do grupo são Visualização da Informação, Realidade Virtual e Aumentada, Interação Humano-Computador, Aprendizado de Máquina, Processamento de Imagens, Visão Computacional, Síntese e Reconhecimento de Voz e Tecnologias Assistivas.

Rua Augusto Corrêa, 1 - Guamá, Belém
Universidade Federal do Pará - Campus Guamá
Instituto de Ciências Exatas e Naturais
Setor Básico
E-mail: ufpalabvis@gmail.com

One of the main challenges to creating rich, seamless, and adaptive Augmented Reality (AR) browsers is the accurate registration of the virtual contents in the real world. Usually, the AR browsers offer augmented navigation functionality by GPS and sensors, such as magnetometer and accelerometer. However, the position of virtual markers suffers some errors when the user is near to the desired location, due to many factors such as sensors failures and bad internet connections, among others. Additionally, to identify the correct marker, when there are many, is a challenging task to users. Therefore, to mitigate these problems, this paper proposes a hybrid approach of location and vision based tracking for AR applications, since the image recognition can be very helpful to identify near locations, avoiding misplaced markers and at the same time giving emphasis to that marker. Furthermore, to avoid bottlenecks in the AR browser applications the combination of the quality of vision-based tracking and the speed of the sensors is proposed. The designed system gets the information about the Points Of Interests (POIs), recommend places to explore around the user via GPS and sensors (as already done by current AR browsers) and run the recognition process only for the nearest POI to improve its registration. Aiming to choose the best recognition algorithm for this scenario, precision and time tests are performed using three algorithms (ORB, BRISK, and AKAZE) to detect keypoints and compute theirs features, and two algorithms (RANSAC and LMEDS) to estimate camera pose. The test pointed that the combination of AKAZE and RANSAC has the best accuracy, but an impractical time to use in real time application. Hence, the usage of vision techniques in an interval of time (skipping some frames) and the usage of inertial sensors movements to update the skipped frames are proposed, in order to use this solution on a mobile platform. Finally, the system solution was implemented in a tourism mobile AR application and some results are presented.

Visual Abstract!

One of the main challenges to creating rich, seamless, and adaptive Augmented Reality (AR) browsers is the accurate registration of the virtual contents in the real world. Usually, the AR browsers offer augmented navigation functionality by GPS and sensors, such as magnetometer and accelerometer. However, the position of virtual markers suffers some errors when the user is near to the desired location, due to many factors such as sensors failures and bad internet connections, among others. Additionally, to identify the correct marker, when there are many, is a challenging task to users. Therefore, to mitigate these problems, this paper proposes a hybrid approach of location and vision based tracking for AR applications, since the image recognition can be very helpful to identify near locations, avoiding misplaced markers and at the same time giving emphasis to that marker. Furthermore, to avoid bottlenecks in the AR browser applications the combination of the quality of vision-based tracking and the speed of the sensors is proposed. The designed system gets the information about the Points Of Interests (POIs), recommend places to explore around the user via GPS and sensors (as already done by current AR browsers) and run the recognition process only for the nearest POI to improve its registration. Aiming to choose the best recognition algorithm for this scenario, precision and time tests are performed using three algorithms (ORB, BRISK, and AKAZE) to detect keypoints and compute theirs features, and two algorithms (RANSAC and LMEDS) to estimate camera pose. The test pointed that the combination of AKAZE and RANSAC has the best accuracy, but an impractical time to use in real time application. Hence, the usage of vision techniques in an interval of time (skipping some frames) and the usage of inertial sensors movements to update the skipped frames are proposed, in order to use this solution on a mobile platform. Finally, the system solution was implemented in a tourism mobile AR application and some results are presented.