Hierarchical real-time recognition of compound objects in images

This dissertation proposes a novel approach for the recognition of compound 2D objects in images under real-time conditions. A compound object consists of a number of rigid object parts that show arbitrary relative movements. The underlying principle of the approach is based on minimizing the overall search effort, and hence the computation time. This is achieved by restricting the search according to the relative movements of the object parts. Minimizing the search effort leads to the use of a hierarchical model: only a selected root object part, which stands at the top of the hierarchy, is searched within the entire search space. In contrast, the remaining parts are searched recursively with respect to each other within very restricted search spaces. By using the hierarchical model, prior knowledge about the spatial relations, i.e., relative movements, between the object parts is exploited already in an early stage of the recognition. Thus, the computation time can be reduced considerably. Another important advantage of the hierarchical model is that it provides an inherent determination of correspondence, i.e., because of the restricted search spaces, ambiguous matches are avoided. Consequently, a complicated and computationally expensive solution of the correspondence problem is not necessary. The approach shows additional remarkable features: it is general with regard to the type of object, it shows a very high robustness, and the compound object is localized with high accuracy even under projective distortions. Furthermore, several instances of the object in the image can be found simultaneously. One substantial concern of this dissertation is to achieve a high degree of automation. Therefore, a method that automatically trains and creates the hierarchical model is proposed. For this, several example images that show the relative movements of the object parts are analyzed. The analysis automatically determines the rigid object parts as well as the spatial relations between the parts. This is very comfortable for the user because a complicated manual description of the compound object is avoided. The obtained hierarchical model is used to recognize the compound object in real-time.