A discrete filter is examined for the joint estimation of the coordinates, velocity, and angular rate W of a target performing a steady turn, based on indirect (polar/spherical) radar measurements. When W is included in the augmented state vector, the increment conversion matrix becomes a nonlinear function of W itself; consequently, linearization – as employed in previously studied versions of the extended Kalman filter for indirect measurements – requires recalculating the Jacobian matrix at each step and introduces an additional methodological error. An unscented transformation of the augmented state vector is proposed to eliminate the need for linearizing the motion and measurement models. Monte Carlo statistical simulations ( N run = 545) were conducted to compare the proposed algorithm with its linearized counterpart across eight standard surveillance radar types (2D and 3D) and five target turn rates. The results demonstrate that the proposed algorithm yields a root-mean-square error for the W estimate that is, on average, 4.3 times lower (median 4.0; range 1.4–14.0 across all observation conditions) than that of the linearized extended Kalman filter, while offering a more modest improvement (averaging 1.2 times) in the root-mean-square error of the position estimate.
This paper presents the results of a statistical synthesis of an algorithm for segmenting images of aerial objects based on the Bayesian criterion of maximum posterior probability. The key feature of the algorithm is the use of information about the operator’s initial choice of the object to form a priori spatial distribution of coordinates, which allows effectively taking into account geometric constraints on the movement of the object between adjacent frames of the video sequence. A twostage approach has been developed to jointly solve the tasks of pixel classification and object position estimation, in which spatial information is directly integrated into the segmentation decision rule through a Gaussian model of probability distribution. Analytical expressions for the optimal decision rule are obtained in the form of a threshold comparison of the log-likelihood ratio, which includes both intensity and spatial components. The resulting algorithm improves the quality of segmentation and the accuracy of coordinate measurements under varying lighting conditions, which is critically important for automatic tracking systems of aerial objects in the tasks of airspace monitoring and flight trajectory management.
The article considers the synthesis of a binary radar recognition device in the absence of a priori analytical models of classification feature distributions. A methodology for sequential radar target recognition is proposed, based on the combined use of support vector machines and Wald’s sequential probability ratio test. At the training stage, a separating hyperplane is constructed, with respect to which the multidimensional feature space is transformed into a one-dimensional decision variable represented by the signed distance to the class-separation boundary. Based on the training sample, the conditional probabilities of correct and false recognition are estimated and used to form increments of the decision statistic. At the recognition stage, the statistic is recursively accumulated and compared with threshold levels until a decision is made. The methodology preserves a unified algorithmic structure when the feature set changes and allows adaptation through repeated classifier training.