Abstract: A new technique for smart city guide using mobile augmented reality is proposed, which satisfies the personalized, multi-scale, comprehensive needs of users and presents active interface with virtual-real fusion. Mobile side is limited by computing power and resource storage capacity. However, mobile devices usually integrate multiple inertial sensors, which are portable and easy to display. Server side is used for city-scale location recognition based on vocabulary tree method. Dynamic partition method with GPS information reduces the range of image retrieval. Hierarchical k-means clustering on BRISK feature with binary descriptors improves the real-time performance of vocabulary tree. Hybrid features based on BRISK and optical flow are executed in parallel for real-time and robust tracking. Regular re-initialization with BRISK feature is used for reducing errors generated by optical flow. Matching point sets mapping is applied for eliminating drift of feature points during initialization of BRISK feature. Sequence frames and keyframe information are used for reducing jitter with pose estimation. Experimental results on UKbench and real environment demonstrate the advantage of virtual-real fusion for city-scale smart guide. Users can easily interact with surrounding real environment. The prototype system has been successfully applied to smart guide system of Shanghai Telecom Experience Venue and other such guide systems.