Research On Correlation Filtering And Tracking Algorithm Based On The Combination Of Convolutional Network Model And Detection Mechanism

Target tracking is widely used in video surveillance,human-computer interaction,visual navigation,medical diagnosis and other fields.In recent years,the technology of target tracking has made significant developments,but it still faces some challenges.The rapid movement of the target leads to the boundary effect,and the detection of the edge is disturbed by the background information,which causes the tracking frame to drift.The tracked target appears in the field of view after being completely occluded for a long time,and the tracking algorithm's wrong response to the background affects the target model update result,resulting in tracking failure.Therefore,this thesis conducts in-depth research on the basis of the tracking algorithm of nuclear correlation filtering,and proposes some improved methods.The main research contents and innovations are as follows:A mechanism for distinguishing and detecting target effective information based on Gaussian mixture model is proposed.First,integrate the local image saliency of the target area to obtain the BRISK algorithm threshold,then sample the key feature points in the image,use the optical flow method to filter out the abnormal feature points of the removed image,and then divide the image in the target area evenly.Block,establish a Gaussian mixture model based on the distribution characteristics of the central feature points of the block image,and then match the remaining feature points in the block image with the model respectively.If the matching is successful,it will be the background feature point,otherwise it will be the target feature point.Gaussian mixture model effectively distinguishes between targets and background information in complex scenarios.Finally,the feature points in the fast image are input into the naive Bayes classifier to calculate the feature score,the position of the output frame is adjusted by the feature score,and the weight value assigned according to the similarity between the block image and the target position forms a strong detector.Furthermore,the influence of background information at the edge of the target in the tracking frame on the classification of the filter is weakened.An improved target tracking method based on multi-convolutional layer feature activation mapping is proposed.First,after image enhancement processing,local features are used to perform non-linear mapping of convolution operations,and the high-resolution network model structure is reconstructed by complementing details.Under the framework of convolutional neural network,multi-scale features are extracted from different resolutions of the image,and then gradients are used.The weighted class activation mapping introduces the spatial gradient of the target pixel into the guided backpropagation,so that the multi-convolution layer feature map highlights the fine-grained details,thereby reducing the error in the target location of the algorithm model caused by the noise in the image.Finally,the weight value of the discriminative sparse coding adaptively fuses the result of the feature response of multiple convolutional layers to improve the accuracy of the tracking algorithm.This thesis compares and analyzes the video sequence experiment with multiple traditional target tracking algorithms.The results show that the target effective information discrimination detection mechanism based on the Gaussian mixture model can alleviate the boundary effect caused by the image blur after fast motion,and reduce the tracking filter's impact on the target.The response is disturbed by background information.When the target appears after being occluded for a long time,the detection mechanism relocates to the target area according to the constraints of the learning module.Improved target tracking method for multi-convolutional layer feature activation mapping,using the detailed features at the edge to obtain complementary high-resolution network models to extract more target feature information,improving the stability of target tracking in complex scenarios,after feature maps are activated and mapped the fusion filter response results,improves the robustness of the algorithm.