With the proliferation of smartphones, recent years have witnessed the rapid development of smartphone-based indoor navigation systems. However, existing solutions either bear high deployment cost or cannot support large-scale navigation. A scalable and plug-and-play indoor navigation system is still highly desirable. In this paper, we propose DeepNav, a new indoor navigation system that fully uses visual CNN to realize large-scale navigation. DeepNav adopts a single-pilot deployment scheme to realize fast deployment. It divides the indoor area into dense sub-areas to simplify image-based location matching while ensuring reasonable resolution. Practical realization of DeepNav entails a set of key challenges, e.g., invalid image recognition, classification of thousands of labels and under-fitting. In order to solve these challenges, we propose invalid image filter, subgroup sigmoid layer and movable object filter, respectively, for DeepNav. Finally, we implement a prototype of DeepNav on commercial smartphones. Experimental results demonstrate that DeepNav can be quickly deployed (e.g., within an hour in a 4-storey building) with an average localization error of 2.3 meters.

随着智能手机的普及，近年来基于智能手机的室内导航系统得到了快速发展。然而，现有的解决方案要么部署成本高，要么无法支持大规模导航。一个可扩展和即插即用的室内导航系统仍然是非常需要的。在本文中，我们提出了 DeepNav，一种新的室内导航系统，它充分利用视觉 CNN 来实现大规模导航。DeepNav 采用单飞行员部署方案进行快速部署。它将室内区域划分为密集的子区域，以简化基于图像的位置匹配，同时确保合理的分辨率。DeepNav 的实际实现需要一系列关键挑战，例如无效图像识别、数千个标签的分类和欠拟合。为了解决这些挑战，我们分别为 DeepNav 提出了无效图像过滤器、子组 sigmoid 层和可移动对象过滤器。最后，我们在商用智能手机上实现了 DeepNav 的原型。实验结果表明，DeepNav 可以快速部署（例如，在一小时内在 4 层楼中部署），平均定位误差为 2.3 米。