Research On Infrared Target Tracking Method Based On Particle Filter

Infrared target tracking is a key technology in the fields of early warning,precision guidance,safety monitoring and so on.However,the infrared image is usually characteristic by low SNR and poor target visibility.Due to the relative motion between infrared imaging system plane and target,the target state and appearance in the field of view keep changing.Within the mainstream particle filter framework,infrared target track method still faces the following problems: it is difficult to predict target state with non-stationary random changing,and the adaptive ability of target appearance model is not enough in different scenes.Besides,search efficiency and location accuracy is usually not high.With the latest research achievements in the frontiers of image processing,pattern recognition and machine learning,In this dissertation,infrared image sequences are employed to research single target tracking method and makes efforts in improving its robustness to interference factors and adaptive ability in different scenes.The main work completed is as follows:When the target state changes suddenly,the coverage efficiency of candidate particle set decreases which is generated by priori knowledge.Thus,it is difficult to predict the target state with non-stationary random changing.To solve this problem,a motion modeling method is proposed for infrared target based on Gaussian sampling and saliency importance sampling.First,target state in the last frame is treated as mean value,and Gaussian sampling is performed to generate particle set which can cover the current target state basically.Then the saliency model is introduced,the image significant region can be obtained when target appearance statistics information is unknown.Thus,the particle set generated by saliency importance sampling can cover salient areas which are nearby target at the last frame.As the experimental results shown,on the premise that the target area is salient compared with its neighbor background,the proposed method can improve the prediction ability for infrared target state non-stationary random changing.To perform multiple features fusion,it is difficult to allocate reasonable weights based on their ability to discriminate target and background,thus the adaptive ability of target appearance model is not enough in different scenes.An infrared target appearance model is proposed which employs multiple kernel fusion based on interframe size invariant.Due to infrared target characteristics of strong maneuverability and poor visibility,the histogram similarity is measured by Bhattacharyya coefficient to construct the kernel function,then the target templates and candidates are mapped nonlinearly from intensity feature space to kernel feature space.The target sizes estimated with different kernel features are compared with that in the last frame,and the kernel feature weights can be allocated based on target size error.As the experimental results shown,on the premise that the target size is interframe invariant,the proposed method can improve the adaptive ability of target appearance model in different scenes.Template matching performed in the kernel feature space is characteristic by low computation efficiency,and errors are introduced by non-effective particles when they are used to estimate target state.It means that the search efficiency and location positioning accuracy are usually not high.A infrared target search mechanism is proposed with template matching based on sparse constraint.First,the target candidates are treated as whole to perform template matching in intensity feature space.The sparse constraint is introduced into the linear regression problem corresponding to the template matching,thus effective subset can be obtained from target candidate set.Then the target candidate effective subset are mapped from intensity feature space to kernel feature space to perform template matching,current target state can be estimated with weighted summation of particle candidate effective subset.As the experimental results shown,the proposed method can improve the search efficiency for infrared target compared with preforming template matching for all the target candidate in kernel feature space.On the premise that target templates and target candidate positive samples are distributed in the same intensity feature linear subspace,the proposed method can improve infrared target locating accuracy.