A Hybrid Localization Framework for Mobile Devices.

We are presenting a novel hybrid mobile localization system for smart mobile devices which combines the prevalence of classical wireless methods with accuracy of computer vision techniques. In the proposed framework, initial localization is roughly measured through classical wireless methods. Later computer vision techniques are deployed to refine the measurement. Specifically an image acquired by the device's camera is matched against 3D model fragments. The key innovation of the framework is to keep a pool of 3D model fragments, indexed by their rough location info, rather than a monolithic 3D model of the environment. Thanks to the location indexed fragments, computer vision computations become reasonable enough to carry on a modern smart-phone. Also, network traffic and mobile storage requirements are minimized by on-demand caching. The 3D models that are needed by the system can come from modern cheap 3D depth sensors or recently matured depth from image sequences techniques. We expect such an approach would be especially useful for Augmented Reality applications that require good localization accuracy.