Research On Extended Target Tracking Technology Based On Feature Representation

With the rise of intelligent devices,the research of computer vision has attracted more and more attention.As one of the research directions,target tracking plays an important role and is widely used in intelligent video surveillance,human-computer interaction,virtual reality,robot visual navigation,video editing and analysis,medical diagnosis and other fields.At present,the research on general target tracking at home and abroad is relatively extensive and has achieved good results,but the research on extended target tracking is very few.The existing algorithms can only solve the tracking problem in specific scenarios.Compared with general targets,the extended target accounts for a larger proportion of the field of view.In addition to Scale Variation,Illumination Variation,Rotation Variation,Deformation,etc.it may exceed the field of view.In addition,the extended target has richer features such as edges,textures,shapes,and key points,which can help the tracking algorithm to build a robust model.Therefore,from the perspective of feature extraction and representation,this paper makes an in-depth study on the extended target tracking algorithm.The main research work of this paper includes:1.Firstly,the research background and significance of target tracking are introduced,and the current research status at home and abroad is analyzed.The discriminative method correlation filtering is introduced.It is found that it has the advantages of stable tracking and can provide a stable framework for extended target tracking.Then introducing the common feature extraction and representation methods,classify them,and analyzing their advantages and disadvantages and use scenarios through experimental comparison.Finally,the generalized Hough transform is studied in depth,and its principle of detecting objects of any shape is analyzed.It is found that its discrete target representation(R-table)has certain guiding significance for the establishment of target models in extended target tracking.2.The extended target tracking scenes are divided into two categories: simple backgrounds and scenes with background interference.Then,for simple backgrounds,an extended target tracking algorithm based on Harris corner and generalized Hough transform is proposed,which can track any selected tracking point.The algorithm uses Harris corners that are easier to extract and more robust to build an R-table table to build the target model and form the target descriptor.Experiments show that the proposed algorithm improves the speed by an order of magnitude while ensuring accuracy,and solves the difficulty of the generalized Hough transform method to accurately extract the edge contour of the target.3.Aiming at scenes with background interference,an extended target tracking algorithm(CF-IGHT)based on correlation filtering and improved generalized Hough transform is proposed.The algorithm first uses the stability of correlation filtering to perform rough positioning to obtain the approximate location of the target,and then uses the accuracy of the improved generalized Hough transform(SURF algorithm for feature point matching)to perform position correction.Through the experimental comparison and Analysis on ten aircraft sequences and OTB2015,it is found that CFIGHT algorithm has achieved good accuracy,and compared with the correlation filter algorithms,such as the Staple,SAMF,DSST,KCF and the generalized Hough transform methods,it has a good improvement,which effectively solves the problem of tracking point offset caused by the correlation filter algorithm due to model interference and the problem that generalized Hough transform methods cannot be applied to scenes with background interference.