Research On Object Tracking Method Based On Correlation Particle Filter

In the field of computer vision,object tracking is the basis of deep analysis for external information by computers.It is also an indispensable application requirement in both national defense and industrial applications.In practice,the complex and changeable environment conditions bring great challenges to object tracking.To improve the robustness of the tracker is the main aim and work of this paper.When estimating the state of moving objects using the particle filter algorithm,the larger the number of particle samples,the more accurate the state estimation.However,the algorithm generates a large amount of calculation,and it easily leads to sample impoverishment in recursive state estimation.This will inevitably result in a drop in its tracking performance.In the kernelized correlation filter algorithm,when the object is occluded,the filtering template is susceptible to contamination,and the position of the object may change quite a bit before and after the occlusion.This might cause the tracker to lose the object.Moreover,when the scale of the object is altered,the fixedsize tracking box cannot adapt to the change,and thus a decrease in the accuracy of tracking ensues.To overcome these three disadvantages,several optimization methods were proposed correspondingly in this paper.Related work is as follows:In order to improve the computational efficiency of the particle filter algorithm and ensure the diversity of particles,this paper proposed a novel particle filter algorithm based on particle proliferation.Firstly,it screened off large-weight particles from the particle set of importance and then generated new particles around these particles through random sampling.The newly obtained particle samples were independent and unique,which ensured the diversity of particles.Besides,there's no need to perform multiple copy operations on large-weight particles.This reduced the calculation amount and improved the efficiency of the algorithm.Finally,through comparative experiments,the effectiveness of the proposed algorithm in improving efficiency and ensuring particle diversity was proved.In order to improve the robustness of the kernelized correlation filter algorithm in object occlusion,this paper proposed a robust occlusion prediction algorithm.Firstly,it divided the object into multiple blocks,used different filtering templates to train them and obtained different classifiers.Secondly,it's combined with the improved particle filter algorithm in the next frame to estimate the state of each block,so as to estimate the object position through the state information of all blocks.Finally,comparison experiments proved that the proposed algorithm was effective in dealing with the object occlusion problem.In order to improve the accuracy of the kernelized correlation filter algorithm when the scale changes,this paper proposed a scale-adaptive method.Firstly,based on the block position,the template was scaled with different coefficients using the size of blocks in previous frames as a fixed template.This process produced multiple scale windows,among which the one with the largest response value after scale filtering was later used as the object area.The corresponding scale factor functioned as the object update ratio.Finally,the paper combined the proposed object occlusion method with this scale update method,and did multiple groups of comparative experiments based on the OTB data.The results showed that the proposed algorithm had a better performance in dealing with object occlusion and scale variation.