Incremental semantic tracking on mobile devices

• Autor(a) principal: ROBERTO, Rafael Alves
• Data de Publicação: 2018
• Tipo de documento: Tese
• Idioma: eng
• Título da fonte: Repositório Institucional da UFPE
• Texto Completo: https://repositorio.ufpe.br/handle/123456789/31917
• topic: Visão computacional; Rastreamento semântico

Tracking is an important task that is used for several applications. The improvement and popularization of mobile devices in recent years allowed these applications to be executed on such devices, which provides a mobility that is not possible on desktop computers. However, there are still several challenges in this field. Thus, the goal of this Ph.D. is to investigate methods to perform tracking on mobile devices considering the characteristics of such platform.To achieve this goal, it was conducted a systematic mapping on tracking for mobile devices. This study collected 2,602 papers, from which 444 were selected to be classified. The results indicated a growing interest in this field and a preference for works that use the device’s sensors to perform tracking locally on the device.The mapping was used to elaborate preliminary experimental scenarios. First, the Google Tango platform was evaluated to establish a ground base of the state-of-the-art trackers. It was observed that the precision in indoor spaces is suitable to provide a good user experience, including for augmented reality applications. Another experiment evaluated the use of parallelism, distributed approach and native implementation. This test showed that, on average, native development was the most efficient. Besides that, experiments were designed intending to test different tracking techniques. One is a face tracking technique using machine learning that was adapted to consider the characteristics of mobile devices and it runs in approximately eight milliseconds on such equipments. The other one is a SLAM technique that was developed in desktop and was ported to a Tango tablet device.There were several lessons learned from the experiments. One of them was the importance of finding high-level semantic information from a scene, which can improve tracking and provide more realistic rendering. In this Ph.D., it was developed a technique that incrementally detects and tracks primitives using the generating process of point clouds of visual SLAM systems, called Geometric and Statistical Incremental Semantic Tracker (GS-IST). The experiment indicates that GS-IST was able to improve both precision and stability of existing methods. However, since it focuses on precision, it compromises the recall to ensure the detection and tracking of correct shapes.In order to evaluate how GS-IST would perform running on mobile devices, it was ported to the Android platform. The evaluation showed that the mobile version is 8.5 to 9.9 times slower in comparison with the desktop implementation. Moreover, it uses up to 30.5% of the CPU load, which allows this implementation to run on a separate thread of the main tracking technique. Additionally, the energy consumption was not a concern because GS-IST can run for more than 4 hours in the worst case. Finally, the memory usage was less than 8% of the total RAM memory of the test devices, which did not have an impact on the execution time.