Research On The Improvement Of GM-PHD And Its Application In The Target Tracking Of Star Background Points

The point target based on space-based platform observation is affected by a large number of interference factors such as stars,clutter,false alarm and noise,which make us have obvious error when using GM-PHD(Gaussian Mixture Probability Hypothesis Density Filter,GM-PHD)to track multiple point targets in the sky background.The main reasons can be summarized as two points: 1)The moving sensor platform creates the background of the moving target,which makes the stars around the target to be tracked have the same trajectory as the moving target.2)The application scenario of GM-PHD algorithm does not consider that the clutter will exist in the observation domain at any time like the target.Because of the above two reasons,the GM-PHD algorithm can not distinguish the observed values of stars and targets.This makes the algorithm take the observation value of the star as the observation value of the real target indiscriminately when performing the update operation.And in the process of the target approaching to the star,the prediction weight of the star will gradually be greater than the 0.5 pruning threshold,and finally the star will be output as the real target derived new target in the process of the target state extraction.Because the main research area of this paper is point target tracking based on space-based platform observation,the tracking environment involved has the above two points.In order to keep the correct tracking of point target in the sky background.In this paper,according to the two different situations of linear motion and nonlinear motion of the target,threshold separation clustering algorithm and motion template weight penalty algorithm are proposed respectively.These two improved algorithms make GM-PHD algorithm have the ability to distinguish stars from real targets,and effectively overcome the tracking difficulties in the starry sky scene.Therefore,the main work and contributions of this paper are as follows:(1)Linear motion of target.To solve the problem of motion sensor platform,we use the sparse target point ICP fast registration algorithm to reduce the impact of star movement.To solve the problem of star time,this paper proposes a simple and effective threshold separation cluster,After pruning and merging in GM-PHD algorithm,the cluster operation is carried out according to the corresponding strategy.The GM-PHD algorithm has the ability to distinguish stars from real targets.For the case that the weight of the target is reduced due to missing inspection.In this paper,a dynamic weight extraction scheme is used to replace the fixed extraction threshold in GM-PHD.(2)Nonlinear motion of target.In this paper,a simple and effective algorithm of joint penalty weight is proposed.The algorithm sets a unique label for each target.For the randomness of the moving direction of the target in the process of moving,we discretize thepossible moving direction of the target into ten irregular motion templates,each partition represents the possible movement direction of the target,and also represents the strength to be punished.Then,we calculate the penalty factor of the target direction and speed by using all the estimates with the same label at k-1 time and the real target at k-1 time,and then calculate the weight matrix and penalty strength matrix after penalty.Finally,we use the new weight matrix to output the target with the largest weight in each tag as the real target in k-time,and remove some fuzzy weights with larger punishment intensity from the punishment intensity matrix through the preset punishment intensity threshold,so as to selectively filter out some targets with less correlation and further avoid Star Information participating in the next iteration.In order to better simulate the scene of star tracking,this paper uses tycho-2 catalog to simulate four kinds of data sets with different complexity for comparative experiments.In order to verify the effectiveness of the two improved algorithms proposed in this paper.In this paper,several tracking algorithms,such as GM-PHD,Forward Backward Smoothing Filter(FBS),N-scan Filter(N-SCAN),Irregular Window Detection Algorithm(GM-PHD-IW),are compared.Experimental results show that the improved algorithm can effectively track multiple space-based platform point targets in the starry sky,no matter the target is in linear or nonlinear motion.