Visual Tracking Algorithm Based On Model Verification

Target tracking is widely used in many fields such as smart home,medical diagnosis,traffic video supervision and so on.With the rapid development of correlation filtering,deep learning theory and computer storage equipment,discriminant tracking algorithm has gradually become the main development trend in the field of target tracking.The discriminant tracking algorithm transforms the target tracking problem into a binary classification problem between the tracking target and the background.Therefore,the algorithm first uses positive and negative samples to train the foreground and background classifiers,and then uses the classifier to find the optimal target region in the subsequent frames.The core of the discriminative tracking algorithm is the three issues of target location,scale and speed.This article combines the mask segmentation model on the online update tracking framework of the discriminative prediction model algorithm,from how to find a more accurate target frame and how to judge the failure of the tracker as well as the improvement in the correct positioning of the target,an algorithm that can be tracked with high precision on the Graphics Processing Unit(GPU)is designed.The research content of the paper is as follows:(1)In order to further improve the size of the predicted target box and the actual target box,a robust model fusion visual tracking algorithm is designed.This algorithm proposes the idea of model fusion and complementarity.First,the target is pre-processed through the multi-scale target search strategy and the fusion context feature strategy,and the dynamic learning rate adjustment strategy is introduced to ensure the accuracy of target positioning,and then the tracking algorithm is combined with the mask.The segmentation model is merged to solve the problem that the target frame and the target are not completely fit,and improve the accuracy of the tracking algorithm.Finally,it was tested on the standard data set VOT2018.The expected average overlap rate is 0.449,the accuracy of the tracking algorithm target frame is effectively improved,and the tracking speed on the GPU is 30 frames per second,meeting the requirement of real-time tracking.(2)In order to improve the overall accuracy of the algorithm,a visual tracking algorithm based on model parallel verification is designed.The model verification strategy is introduced on the basis of the model fusion algorithm,and the average peak energy ratio is used to judge the reliability of the online tracking algorithm.When an unreliable tracking result appears,the algorithm of non-online update classifier is started to relocate the target position,and finally the model verification is performed.Optimal solution,thereby improving the robustness of the tracking algorithm.Finally,it was tested on the standard data set VOT2018.The expected average overlap rate is 0.453,this further improves the accuracy and stability of the algorithm.The tracking speed on the GPU is 32 frames per second,meeting the requirement of real-time tracking.