Date of Degree
Artificial Intelligence and Robotics | Computer Engineering | Computer Sciences
assistive indoor localization, real-time system, GPU acceleration, mobile computing, omnidirectional vision, geometric computing, structure from motion, RGB-D
An indoor localization system is of significant importance to the visually impaired in their daily lives by helping them localize themselves and further navigate an indoor environment. In this thesis, a vision-based indoor localization solution is proposed and studied with algorithms and their implementations by maximizing the usage of the visual information surrounding the users for an optimal localization from multiple stages. The contributions of the work include the following: (1) Novel combinations of a daily-used smart phone with a low-cost lens (GoPano) are used to provide an economic, portable, and robust indoor localization service for visually impaired people. (2) New omnidirectional features (omni-features) extracted from 360 degrees field-of-view images are proposed to represent visual landmarks of indoor positions, and then used as on-line query keys when a user asks for localization services. (3) A scalable and light-weight computation and storage solution is implemented by transferring big database storage and computational heavy querying procedure to the cloud. (4) Real-time query performance of 14 fps is achieved with a Wi-Fi connection by identifying and implementing both data and task parallelism using many-core NVIDIA GPUs. (5) Rene localization via 2D-to-3D and 3D-to-3D geometric matching and automatic path planning for efficient environmental modeling by utilizing architecture AutoCAD floor plans.
This dissertation first provides a description of assistive indoor localization problem with its detailed connotations as well as overall methodology. Then related work in indoor localization and automatic path planning for environmental modeling is surveyed. After that, the framework of omnidirectional-vision-based indoor assistive localization is introduced. This is followed by multiple refine localization strategies such as 2D-to-3D and 3D-to-3D geometric matching approaches. Finally, conclusions and a few promising future research directions are provided.
Hu, Feng, "Vision-based Assistive Indoor Localization" (2018). CUNY Academic Works.