Real-Time Tracking of Complex Objects Using Dynamic Interpretation Tree

Abstract

Vision-based tracking for augmented reality (AR) applications requires highly accurate position and pose measurements at video frame rate. Typically several interaction devices have to be tracked simultaneously. While the geometry of all devices and the spatial layout of visual landmarks on the devices are well known, problems of occlusion as well as of prohibitively large search spaces remain to be solved. The main contribution of the paper is in high-level algorithms for real-time tracking. We describe a model-based tracking system which implements a dynamic extension of the structure of an interpretation tree for scene analysis. This structure is well suited to track multiple rigid objects in a dynamic environment. Independent of the class of low-level features being tracked, the algorithm is capable to handle occlusions due to a model-dependent recovery strategy. The proposed high-level algorithm has been applied to stereo-based outside-in optical tracking for AR. The results show the ability of the dynamic interpretation tree to cope with partial or full object occlusion and to deliver the required object pose parameters at a rate of 30 Hz.