Research On Object Tracking Algorithm Based On Dynamic Feature Fusion And Model Update

With the rise of artificial intelligence boom,computer vision has been developed rapidly in recent years.As an important part of computer vision,visual object tracking has received widespread attention and in-depth research.Although researchers have put forward many effective object tracking algorithms in recent years,there are still many challenges in practical applications.In particular,factors such as illumination change,deformation,occlusion,and background clutter in the tracking process have a great influence on the performance of the object tracking algorithm.Because of the limitation of single feature,tracking algorithms are often difficult to deal with all kinds of interference during tracking.Based on the current mainstream target tracking algorithm framework,this paper improves the feature fusion method,model updating strategy and scale estimation in the tracking algorithm,then proposes an improved tracking algorithm.The main contents and results of this paper are as follows:4.The existing two mainstream multi-features fusion algorithms SAMF and Staple are analyzed,and the defects of their fusion methods are pointed out.By analyzing the reliability of response diagram,a method of target response fusion based on dynamic weight is proposed.By combining the three features of HOG,CN and color histogram,and the two tracking frames with correlation filter tracking and color probability tracking,a dynamic feature fusion strategy of three layers fusion is designed.5.Analyze the situation that the introduction of noise during the update of the model leads to the failure of the model and eventually results in the failure of tracking,introduce the traditional model update method and its defects.Using the interference detection to propose a method for calculating the update rate of the model and implementing a dynamic model updating strategy.Finally,a new tracking algorithm DMUCT is proposed by combining dynamic feature fusion with model updating strategy.Through experiments,the accuracy of the DMUCT algorithm compared to Staple and SAMF in the OTB2013 standard data set has improved by 4.8% and 5.2% respectively,and the success rate has increased by 6% and 3.5% respectively.6.Combining the probability segmentation in the DAT algorithm and the multi-scale response estimation in the SAMF algorithm,combining the two algorithms in a cascaded way to improve the scale estimation algorithm.Firstly,the first scale estimation is performed using the probability map segmentation method.Then,based on the results of the first scale estimation,it is decided whether or not to use the multi-scale response method to perform the second scale estimation.Applying the improved strategy to the DMUCT algorithm to obtain a new DMUCT-S algorithm.Through experiments,it is found that the improved scale estimation algorithm significantly improves the accuracy of the target scale in the target tracking process,and also improves the overall performance of the target tracking algorithm.At the same time,this method is used to improve other objecttracking algorithms.Compared with the tracking results,it also finds that the method has significant improvement in the scale estimation effect of other tracking algorithms.