Outdoor Target Recognition And Its Applications In Augmented Reality

The application of Augmented Reality (AR) technologies in the large scale is one of the hot topics in the community. Target recognition is one of the key technologies in the task of the application, because the computer needs to locate and recognize target of interest in an image. However the various categories of targets and noises cause the task extremely difficult.Based on the different types of targets, the thesis divides the task of target recognition into two categories: (1) One is the scene recognition, in which the whole image is a target; (2) and the other is object detection and recognition in which an image contains multiple targets. Although the computer vision provides rich approaches to the task of scene recognition and the task of object detection and recognition, these methods rarely take the real-time performance into account. Most scene recognition algorithms use solo type of feature and cause the failure of classification between similar scenes. Besides, when the number of scenes increases, the efficiency and accuracy of algorithms drop dramatically. In the object detection and recognition task, part based model has good performance in the term of detection, but performs poor in the term of recognition.To improve the accuracy and efficiency of the target recognition task in AR, the thesis mainly focuses on the task of scene recognition and the task of object detection and recognition. The work of this thesis is as follows,(1) To solve the problem of classification between similar scenes, the thesis proposes a hybrid feature approach which uses both GPS and visual information. In the recognition stage, the algorithm firstly finds the use's corresponding region to coarsely locate the scene. Then the proposed hybrid feature model is used to accurately locate the scene by combining texture, contour and color features. The results illustrate the algorithm both robust and efficient.(2) To solve the problem of poor efficiency and accuracy caused by enormous scenes, the thesis presents a geo-based search tree to hierarchically divide scenes automatically. A geo-based search tree with branching factor K is constructed by recursively using cluster method to split scenes. The atomic region is defined on the leaf node of the search tree. Since the number of scenes in each atomic region keeps relatively small, the search procedure is very fast and recognition rate can be improved.(3) To solve the problem of modeling object configurations, the thesis proposes a reconfigurable template for object detection and recognition. The template can learn part configurations that capture the spatial correlation of features. To describe parts, the thesis defines a dictionary of rectangular primitives of various sizes, aspect-ratios and positions. The optimal configuration is a subset of non-overlapping primitives from this dictionary learned by And-Or search. The optimal configuration can improve performance in terms of both detection and recognition.The results of the experiment section illustrate the algorithms of two tasks presented in this thesis can achieve excellent performance to solve target recognition problem in the complicated outdoor environment.