FDSST Infrared Target Tracking Method Based On Spatiotemporal Context Similarity

In today's world,human beings are in the information age,computer vision is more and more widely used in various fields of life.Infrared image has the characteristics of low resolution and clarity while giving full play to its advantages,which makes target tracking in infrared scene become a research difficulty.Aiming at the problem that traditional methods can not extract effective features to express the target,this paper takes FDSST(Fast Discriminatory Scale Space Tracking)and STC(Spatio Temporal Context)as the framework to study.In FDSST tracking algorithm,two independent correlation filters are designed to estimate the position and scale of the target of interest respectively.To track the target only by using the local features of the target,this paper proposes to establish a similarity model to compare the similarity of the location information obtained by the two tracking algorithms,and uses the spatio-temporal context around the target in STC tracking algorithm to assist the supervision of the local features of the target,so as to integrate the advantages of the two algorithms.Secondly,for the problem of scale inaccuracy in STC tracking algorithm,we no longer use the update strategy of STC tracking algorithm,but use the result of scale filter to update the scale.After the similarity model is established,the similarity model is further optimized based on the similarity model,aiming at the problems that it can not effectively determine the tracking state and the target can not be relocated.Firstly,the preliminary judgment is made by similarity,and then the secondary judgment is made by average peak to correlation energy(Average Peak-to-Correlation Energy,APCE).Secondly,the updating of the model is stopped in time after the tracking failure is determined,so as to avoid integrating the wrong samples and ensure the reliability of the samples.Thirdly,for the failed target,this paper proposes the improved optical flow method based on SUSAN(Smallest Univalue Segment Assimilating Nucleus,SUSAN)corner and block hierarchical clustering to relocate the target and realize the continuous tracking of the target.Finally,for the infrared scene with low background difference,because the original algorithm can not track normally,this paper proposes an improved Vi Be(Visual Background Extractor,Vi Be)target detection algorithm to separate and track the target background,so as to improve the adaptability of the original algorithm for the infrared scene.By comparing the algorithm in otcbvs,LTIR and other infrared databases,the results show that the algorithm can adapt to more scenes than the original algorithm,and the robustness is improved.