Study Of Multi-bernoulli Filter And Its Applications To Track Before Detection

AbstractWhether in the military or in civilian areas, multiple target tracking techniqueshave shown a wide range of applications. Multi-target tracking algorithms based onrandom sets are research focus in recent years, which are also the research emphases ofthis paper.Dim small target detection and tracking is the key technology of the infraredsurveillance system and precise guidance system. How to detect and track the dimsmall targets in low SNR images is a very challenging research topic.This paper focuses on the theory and application of PHD and the Multi-Bernoullifilter, and the algorithms of small targets track-before-detect in low SNR images.The article first describes the theory of random sets, then studies the algorithm ofGaussian mixture PHD and PHD smoothing algorithm. Simulation results show thatGaussian mixture PHD smoothing algorithm can obtain more accurate state estimationthan the Gaussian mixture PHD algorithm.Second, we do research on the emerging multi-Bernoulli filter, Then itsimplementation--the Gaussian mixture Cardinality Balanced multi-Bernoulli filter isintroduced under linear Gaussian conditions. Then in the non-linear conditions twokinds of Gaussian filtering method-Quadrature Kalman filter and Cubature Kalmanfilter are studied. And the Cardinality Balanced multi-Bernoulli algorithm based onQuadrature Kalman filter and the Cardinality Balanced multi-Bernoulli algorithmbased on Cubature Kalman filter are proposed. They solve the nonlinear filteringproblem very well.Lastly, aiming at solving huge amount of storage and poor tracking results ofmulti-TBD algorithm based on the particle filter, a multi-Bernoulli TBD algorithmbased on the Gaussian-particle filter is proposed. At the prediction and update stages,first some particles are sampled from the Gaussian term, and then the mean andcovariance of the Gaussian term are updated. In each frame we just need to store themean and covariance of the Gaussian distribution, which greatly reduces the amount ofstorage. Even in the case of sampling very few particles, the proposed algorithm alsoreceives a high tracking accuracy.