Research On Infrared Small Target Tracking Algorithm Based On Random Finite

Infrared small target tracking is a challenging image processing task that requires accurate detection and tracking of targets in noisy and interfering backgrounds.In target tracking algorithms,Random Finite Set(RFS)is a mathematical tool used to describe the state of multiple targets.It can handle the appearance and disappearance of targets and the uncertainty of the number of targets.Due to the excellent mathematical properties of RFS in dealing with target tracking problems,infrared weak small target tracking technology based on RFS framework has attracted widespread research by scholars at home and abroad in recent years.However,the problems faced by infrared target tracking under the RFS framework are mainly reflected in: 1.Infrared video tracking scenes may have target scale changes,occlusion,non-planar rotation,background clutter,etc.,which bring difficulties to target detection and feature description.2.Video tracking algorithms need to consider how to choose appropriate observation models and likelihood functions,how to adaptively integrate multiple feature information,how to effectively manage and update model parameters,etc.3.Prior parameters usually need to be set or estimated according to specific scenes and data,but in actual situations,these parameters are often unknown or difficult to obtain.If the prior parameters are set unreasonably or inaccurately,it will cause errors or instability in the tracking results.Based on the above problems,this thesis makes the following research achievements on infrared weak small target tracking under the RFS framework:1.Establish an integrated framework for infrared weak small target detection: In the image preprocessing stage,a multi-frame accumulation bad point removal algorithm is proposed.The bad point position is extracted and the bad point is cleared by solving the gradient image weighted average response in four directions on the difference image between the multi-frame cumulative image and the median filter.In the detection stage:according to the image features and tracking scene complexity,this thesis adaptively selects the detector.First,use the information entropy and discrete coefficient of the image to classify the picture by weighting,and then use the Histogram of Oriented Gradient(HOG)feature saliency algorithm and small bounding box filtering algorithm to detect the weak small target position according to the complexity of the image.The proposed algorithm can effectively screen out bad points in the scene during preprocessing and adaptively select corresponding algorithms based on scene information during tracking.2.Based on Generalized Labeled Multi-Bernoulli(GLMB)filtering,an adaptive birth algorithm suitable for video scenes is proposed.For the target detection probability of position in the tracking scene,based on the Bootstrap idea,the Cardinality Probability Hypothesis Density(CPHD)is integrated into the GLMB filter to jointly estimate the unknown detection probability and weak targets in the scene.Finally,these two algorithms are combined to propose an adaptive GLMB filtering algorithm.The proposed algorithm can effectively track infrared weak targets in different scenes and is more robust than traditional adaptive birth algorithms.3.An improved Probability Hypothesis Density(PHD)filtering algorithm based on Variational Bayesian(VB)is proposed.The algorithm establishes target models and measurement models based on multiple features in the image and proposes a multi-feature likelihood function for updating multi-target states.For unknown measurement noise in the scene,VB inference is used to construct joint posterior intensity of multiple targets to jointly estimate unknown measurement noise and target state in the scene.For the problem of unstable potential estimation of PHD filtering under high clutter conditions,the Hungarian algorithm and M/N track initiation method are combined to propose a track fragment removal method to improve the accuracy of PHD filtering potential estimation.The proposed algorithm can stably track infrared weak targets when the prior knowledge of filtering is unknown.