Multi-Scale Modeling Of 3D Objects And Recognition Method Using Mixed Neural Networks

Method based on models is widely used in 3D target recognition. It needs to get models first. Avoiding the recovery of 3D object from images, Multi-view modeling is expected promising in 3D object modeling. This thesis presents a multi-view and multi-viewpoints characteristic views modeling method and discussed the recognition algorithm of moving 3D objects based on the models.Basic theory and method of 3D object simulating and multi-view model is introduced such as divide of viewing space, characteristic views. (1) The side-glance image sets of 3D airplane objects were obtained after the viewing space was divided uniformity by MultiGen and Vega, the famous 3D simulating software. Subspaces whose feature vectors are similar can be combined. After combination each new subspace has a characteristic view by target projecting to the space. The combination complete by clustering. (2) And the necessity of establishing the multi-scale and multi-viewpoints characteristic view models of 3D target, and the rationality of using the general constraint of target's moving character for recognizing target's image sequence were discussed. (3) A single frame recognition algorithm was proposed based on multi-scale models. And a mixed neural networks multi-frames recognition algorithm was constructed based on multi-scale models, which the multi-scale BP networks was used to classify object, the RBF networks was used to recognize the object poses. The algorithm used reasonable restrictions of the influence to recognition by object scales and the object poses which were not change acutely, and improves the recognition ratio. The training of our algorithm is easy, and needs less samples composed of the object characteristic views models. The single frame recognition and multi-frames recognition algorithm based on multi-scale models can treat with single frame image and image sequence effectively. (4) The rationality and validity of the approach are proved by the results of massive simulation experiments and the results of real video of objects recognition on several kinds of aircrafts.