Neural Computational Modeling For Moving Object Extraction

Moving object extraction is not only the most important research direction in the computer vision field, but also the critical task of location, extraction, tracking and description of the moving object from a video sequence. Although many effective methods have been proposed by researchers so far, most of them need to learn a background model before extracting. However, background modeling methods usually are not satisfied due to the complexity of the background, different types and varying sizes of the targets. In order to tackle with these problems, this thesis establishes a neural computational model for moving object extraction without background modeling.The main contributions are as follows:1. The principle and structure of ART2-A neural network are studied first, and the influence of the parameters of ART2-A is analyzed. A new concept of adjusting vector is introduced to overcome indistinguishable colors information, which offers a good classification accuracy of color pixels and lays a good foundation for the later moving target detection.2. An ART2-A model based on frame difference method is proposed for dealing with the problem of large calculation of ART model. In order to reduce the number of pixels on video,only the pixels of moving regions detected by frame difference are put into the ART2-A neural network, which can increase the computational efficiency.3. A neural computational model based on adaptive resonance theory is introduced to extract moving object. The pixels with high repeated neuron activation are eliminated through the adaptive clustering of ART2-A. Meanwhile the background image is reconstructed through storing the pixels with high repeated neuron activation. Experimental results demonstrate that the proposed algorithm is effective and efficient in moving object extraction without background modeling.