The Research Of Maneuvering Target Tracking Based On Wavelet Neural Network

Maneuvering target tracking is a hot and difficult point on the international nowadays, it makes use of moving target measurement which gains from probe to estimate the target state, it is also a process of removing error, because measurement data contains large amount of disturbance which must be removed. These years, many algorithms have been used in maneuvering target tracking, mass of them are based on target dynamic modeling technology, and many dynamic models have been advanced. However, the main difficulty of maneuvering target dynamic modeling is the uncertainty of target dynamic model, thus superior adaptive method need to be considered, the method is selecting the appropriate model which is most close to target real dynamic state. In the view of the drawback of target tracking based on neural network, the paper studies a modified algorithm of adaptive tracking of maneuvering target.Firstly, this Paper analysis the function and characteristic of wavelet transform and neural network. Wavelet transform has excellent abilities of the rapid convergence rate and approximation accuracy. The neural network is known as ability of self-learning, self-organizing self-adapting and nonlinear mapping. Wavelet Neural Network (WNN) is such a network that its hidden activation function is wavelet, WNN has both advantages of wavelet transform and neural network. This paper applies appearance, research process and application field of WNN, advances a network of multi-resolution orthogonal WNN, such network conquers BP network disadvantages of converge to minimum of one part, and has no redundancy which is usually appears in continuous WNN. Simulation results show that multi-resolution orthogonal WNN has better prediction performance than BP neural network.Secondly, based on basic theory and method of target tracking, the paper studies maneuvering target tracking algorithm based on dynamic modeling. For the realization of kalman filtering algorithm, spherical coordinates are converted to orthogonal coordinates, then measurements are computed in filter. Several maneuvering model and their modeling fashion are compared, they are white noise acceleration model (Constant Velocity-CV), Wiener acceleration model (Constant Acceleration-CA), Singer acceleration model and Current Statistical model-CS. Monte carlo simulation is used in all the target tracking simulation of every model, the simulation results show their tracking accuracy, advantages and disadvantages, and indicates that Current Statistical model is the appropriate model which more close to target's real dynamic state.Problems of tracking accuracy and robust of non-maneuvering target or weak maneuvering target have not been resolved perfectly. Error Back Propagation (BP) network combines with Current Statistical model, double filters structure is used to construct target tracking system. The system can quickly respond to target strong maneuvering, it also can respond to target weak maneuvering adaptively. The tracking structure in a certain extent improves tracking accuracy of non-maneuvering or weak maneuvering target, but the paper proposes a new Current Statistical model target tracking system based on multi-resolution orthogonal wavelet neural network, the simulation results show that the system has faster rapid convergence rate and higher tracking accuracy, and it is more robust.