Research On Single Target Tracking Algorithm Of Siamese Network Based On Correlation Filter

In modern application engineering,Computer Vision and Artificial Intelligence have become important research contents,in which Object Tracking has important application value in solving problems such as automatic driving,monitoring safety,human-computer interaction,and intelligent transportation.Object Tracking is an important issue in the field of computer vision.Through the given object size and position information,the moving target is found in the subsequent image sequence.In practical applications,while improving the efficiency of object tracking,it is necessary to solve complex problems such as target appearance chan ges,scene changes,similar object occlusion,and high-speed motion blur.This paper mainly studies the object tracking algorithm of siamese network based on correlation filter.First,using the loss function of the spatial regularization constraint with an inverse Gaussian function,the reverse gradient transfer adaptively changes the initialized cosine window to obtain a cosine window more suitable for the model.The model fixes the cosine window parameters during the tracking process,does not affect the tracking speed,and improves the tracking accuracy success plot while achieving real-time performance.So far,it is the first single-object tracking algorithm implemented,which uses the same training set as the related algorithm and obtains a better tracking effect on the video test set.Although the above object tracking algorithm can achieve real-time and better tracking effect on the video of general scenes,the model update adopts the strategy of retaining new samples and discarding old samples,which is prone to drift when the tracking object is occluded and lost,Make the learned filter contaminated by background and obstruction.Therefore,in addition to adding certain constraints in space,this paper also studies model updating strategies in time series.In this paper,the Gaussian mixture model is used in the tracking stage to assign historical frames to the corresponding slots according to the feature similarity,so as to achieve the distribution of high similarity within the group and the large difference between the groups,so as to obtain a variety of templates.This method makes the original algorithm improve the tracking accuracy and accuracy of the original algorithm in the standard library while keeping the tracking speed not greatly affec ted.In addition to grouping and updating the model in the tracking phase,this paper also uses the gate signal to control the storage and update mechanism of the model grouping in the training phase,so that the template has diversity and the features obtained by the convolutional neural network are more discriminative,thereby improving the tracker.Accuracy.The template storage and update in the training phase use gate signals to control the position of the slot and the fusion method corresponding to different input features.This part studies the template update method from the two stages of training and tracking.It has good tracking results in the video standard library,and it has better performance in the specific scenes where long videos and tracking targets have obvious feature changes.Finally,this article combines the above research content to implement a single object tracking system.The system can select the tracking model under different methods.After the target calibration of the test video,the target tracking results are displayed in real time.At the same time,the appropriate parameters are adjusted after the different methods to achieve better tracking effect.