In this paper, a geometry-based point cloud reduction method is proposed, and a real-time mobile augmented reality system is explored for applications in urban environments. We formulate a new objective function which combines the point reconstruction errors and constraints on spatial point distribution. Based on this formulation, a mixed integer programming scheme is utilized to solve the points reduction problem. The mobile augmented reality system explored in this paper is composed of offline and online stage. At the offline stage, we build up localization database using structure from motion and compress the point cloud by the proposed point cloud reduction method. While at the online stage, we compute the camera pose in real time by combining an image based localization algorithm and a continuous pose tracking algorithm. Experimental results on benchmark and real data show that, compared with the existing methods, this geometry-based point cloud reduction method selects a point cloud subset which helps the image-based localization method to achieve higher success rate. Also, the experiments conducted on a mobile platform show that the reduced point cloud not only reduces the time consuming for the initialization and re-initialization, but also makes the memory footprint small, resulting a scalable and real-time mobile augmented reality system.