Research On Implementation Of Non-linear Filtering Algorithm Based On FPGA

With the development of technology, target tracking has got more and more widely application in military and civilian field. Target tracking system's performance is affected by the number of objectives, density and dynamic performance etc. Under complex conditions, the nonlinear relationship between sensor measurements and target state makes the nonlinear filtering the difficult and hot issues.However, the nonlinear filtering algorithms with high precision often have a large amount of calculation and a poor real-time.FPGA has a strong parallel computing capability, making full use of it will greatly improve the computing speed of the nonlinear filtering algorithms.Aming to the problem above, this paper focuses on the research on FPGA implementation of Quasi-Monte-Carlo Gaussian Particle Filter (QMC-GPF) and Unscented Kalman Filter (UKF).Based on the analysis of the QMC-GPF principle and structure, a parallel processing structure is proposed: parallel QMC module and parallel GPF module. Base 2 is used to generate Faure sequences, instead of multiplication and mod only bitwise XOR, which is easily to realize on FPGA, is needed to generate the sequences. Look-up tables are used in calculating the complex functions such as exponential function, which make full use of the large number of Block RAM of FPGA. The parallel structure is designed to compute the elements of Cholesky decomposition matrix. DSP48Es are used to realize floating point multiplication and addition, which improve the precision and speed. Infrared imaging dim small target tracking is realized on FPGA and the results show the efficiency and real time of the design.Based on the analysis of the UKF principle and structure, a parallel processing structure is proposed: parallel Unscented Transformation module and parallel Kalman Filter module.The Cholesky decomposition of the diagonal block matrix is converted into sub-matrix on the diagonal Cholesky decomposition, reducing the dimension of decomposited matrix.The matrix inversion is realized by singular value decomposition and the matrix is divided into multiple 2×2 sub-matrix parallel processing units. The Xilinx CORDIC IP cores with pipeline structure are used to calculate the rotation angle and the trigonometric functions, which improves the precision and speed. The three stationary passive sensors tracking a single moving target on the same plane is realized on FPGA and the results show the efficiency and real time of the design.