Systems with a single sensor (radar or infrared image sensor) have their limitations in target recognition and tracking. A system with multi-sensors can fuse data from different sensors to overcome the limitations in the system with a single sensor, it can make use of the complement and redundancy of data from different sensors to improve the precision of target recognition and tracking and the robustness and reliability. In our system for target recognition and tracking, radar and infrared image sensors are used. As a radar sensor in our system can provide with the information of the distance and direction of the target (not the image of the target), data fusion is implemented only at characteristic level and at decision level. For data fusion at characteristic level, characteristics of a target obtained from radar can be used in the IR Image-based subsystem to improve the ability of object recognition, and vice versa. The process of target recognition based on IR image analysis is composed of image enhancement, image segmentation and recognition of segmented objects. For image enhancement, median-filter, histogram equalization, wavelet transformation and canny-operator are used. A median-filter is used for the elimination of punctate noises in an IR image. The grayness of a pixel in an IR image is determined by the grayness of its neighbors. Histogram equalization is used for the improvement of the contrast of an IR image. Multi-scale pyramidal wavelet transformation is used to delete unexpected edges and improve the continuity of edges in an IR image. Canny operator based on Gauss-function is used for edge detection, an IR image is transformed into a binary image based on the adaptive threshold. According to area (number of pixels) of segmented objects, the recognition of segmented objects is divided into two classes: recognition of dot targets and area targets. Rule-based reasoning is used to deal with the recognition of dot targets; a classifier based on neural network is used to deal with the recognition of area targets. The models for target recognition are extracted by data mining. The rules for the recognition of dot targets are extracted by decision trees. A neural network for the recognition of area targets is constructed by multi-layer preceptron and trained by training examples. The following characteristics of target are used as inputs of the neural classifier: •distance of the target obtained from the radar-based subsystem. •area of the target in the IR image, the variation of areas of the target in the consecutive two IR images. •the mean grayness of pixels of the target. •the variation of centers of the target in the consecutive two IR images. •the topological shape of the target (seven moment invariants Ψ1,Ψ2,...,Ψ7, the number of forks in the frame extracted.) •the direction of the target motion predicated by radar and the relation of angles among the axes of the missile, Radar and IR image sensor. After data fusion at characteristic level, a true target is recognized by the radar-based subsystem and the IR image-based subsystem. Based on these two decisions, data fusion at decision level is to make a final decision of target tracking. a factor "decision certainty" is introduced to realize data fusion at decision level, which represents the relative certainty of decisions of target tracking.

A software tool for data mining (DMiner-I) is introduced, which integrates pattern recognition (PCA, Fisher, clustering, HyperEnvelop, regression), artificial intelligence (knowledge representation, decision trees), statistical learning (rough set, support vector machine), computational intelligence (neural network, genetic algorithm, fuzzy systems). It consists of nine function models: pattern recognition, decision trees, association rule, fuzzy rule, neural network, genetic algorithm, HyperEnvelop, support vector machine, visualization. The principle, algorithms and knowledge representation of some function models of data mining are described. Nonmontony in data mining is dealt with by concept hierarchy and layered mining. The software tool of data mining is realized by Visual C++ under Windows 2000. The software tool of data mining has been satisfactorily applied in the prediction of regularities of the formation of ternary intermetallic compounds in alloy systems, and diagnosis of brain glioma.

Wavelet Domain Hidden Markov Model (WD HMM), in particular Hidden Markov Tree (HMT), has recently been proposed and applied to gray level image analysis. In this paper, color texture analysis using WD HMM is studied. In order to combine color and texture information to one single model, we extend WD HMM by grouping the wavelet coefficients from different color planes to one vector. The grouping way is chose according to a tradeoff between computation complexity and effectiveness. Besides, we propose Multivariate Gaussian Mixture Model (MGMM) to approximate the marginal distribution of wavelet coefficient vectors and to capture the interactions of different color planes. By employing our proposed approach, we can improve the performance of WD HMM on color texture classification. The experiment shows that our proposed WD HMM provides a 98% ercentage of correct classifications (PCC) on 44 color images from an Oulu Texture Database and outperforms other methods.

In this paper, we present illumination-affine invariant methods based on affine moment normalization techniques, Zernike moments, and multiband correlation functions. The methods are suitable for the illumination invariant recognition of 3D color texture. Complex valued moments (i.e., Zernike moments) and affine moment normalization are used in the derivation of illumination affine invariants where the real valued affine moment invariants fail to provide affine invariants that are independent of illumination changes. Three different moment normalization methods have been used, two of which are based on affine moment normalization technique and the third is based on reducing the affine transformation to a Euclidian transform. It is shown that for a change of illumination and orientation, the affinely normalized Zernike moment matrices are related by a linear transform. Experimental results are obtained in two directions; the first is used with textures of outdoor scenes while the second test is performed on the well-known CUReT texture database. Both tests show high recognition efficiency of the proposed recognition methods.

可视跟踪就是利用图像处理、模式识别的方法发现视频序列中与指定目标图像最相似的部分，在兼顾实时性的基础上提高跟踪算法的稳健性一直是可视跟踪研究中的前沿和热点。本文提出利用目标历史模型和当前匹配位置处得到的观测模型对目标核函数直方图进行Kalman滤波，从而对模型进行及时更新。首次提出把滤波残差作为样本进行假设检验，将其结果作为模型是否需要更新的准则。论证了Mean-shift框架下跟踪变尺度目标的充分条件，提出了“后向跟踪－形心配准”的核窗宽自动选取算法。实验验证了所提方法的有效性。