Appearance-Based Statistical Object Recognition Including Color and Context Modeling

Within the scope of this dissertation a system for appearance-based statistical classification and localization of 3-D objects in 2-D digital images is presented. The first three chapters define the object recognition task, present the mathematical background of the system, and discuss the known approaches for object recognition. Then the learning phase of the system is described. The training begins with the image acquisition, which can be done with a hand-held camera. The object...
More Description

Within the scope of this dissertation a system for appearance-based statistical classification and localization of 3-D objects in 2-D digital images is presented. The first three chapters define the object recognition task, present the mathematical background of the system, and discuss the known approaches for object recognition. Then the learning phase of the system is described. The training begins with the image acquisition, which can be done with a hand-held camera. The object poses required for the further object modeling are computed from the training image sequence with the structure-from-motion algorithm. In contrast to shape-based approaches, appearance-based methods do not use any segmentation steps to extract object features. The objects are described by 2-D local feature vectors computed directly from image pixel values using the wavelet transform. Both gray level and color images can be used for feature extraction. Finally, the object features are statistically modeled with the normal distribution and stored in the object models as density functions. Additionally, context modeling is also performed in the training phase. In the recognition phase the system classifies and localizes objects in scenes with real heterogeneous background, whereas the number of objects in a scene is unknown. First, feature vectors are calculated in the scene with the same method as in the training. Second, a maximization algorithm evaluates the learned density functions with the extracted feature vectors and yields classes and poses of objects found in the scene. Experiments made on a real data set with more than 40000 images compare the classification and localization rates for all algorithms discussed in the dissertation and show a very good performance of the system in a real world environment.