Research On Planar Object Tracking Algorithm Based On Deep Homography Matrix Estimation

With the popularization of digital media technology and the rapid development of computer vision,planar object tracking as an important 3D tracking technology has been widely used in 3D reconstruction,military guidance and unmanned aerial vehicles and other technical fields.Although many researchers have achieved fruitful research results in this field,there are still many unsolved problems.For example,in scenes with complex backgrounds or fastmoving viewpoints,repeated or missing textures in the video image will cause the extracted key point feature information to be inaccurate;scenes beyond the field of view or occlusion can easily cause the region-based method to fail;the viewpoint moves quickly It will cause a greater degree of deformation of the tracking target,resulting in an increase in the error of the map in the matching process.Existing algorithms are still far from meeting the accuracy and robustness requirements in complex actual scenarios.Traditional planar object tracking algorithms are mainly divided into three categories: keypoint-based,area-based and graph-based methods,but they all have certain limitations.Most keypoint-based methods ignore the structural information of the planar target and rely heavily on the gradient information of the pixels in the image;the region-based method may have tracking failure when the target structure information is lacking;the map-based method is used in texture-rich In the scene,the deviation of feature point extraction will affect the creation of local features and make the tracking not stable enough.Since deep learning can extract target depth features more accurately,it has been widely used in target detection,recognition,and tracking tasks,which significantly improves the accuracy of detection,recognition,and tracking.However,there are relatively few studies on depth-based planar object tracking.Based on this,this paper proposes a planar object tracking algorithm using convolutional neural network to estimate the homography matrix,and on this basis,proposes an improved algorithm for homography matrix estimation based on the siamese neural network.The main contents include:(1)Thesis proposes a planar object tracking method that uses convolutional neural networks to extract features and perform regression estimation on the homography matrix.In order to reduce the difficulty of directly estimating the homography matrix,the algorithm reduces the degree of freedom of the homography matrix through normalization,and uses fourpoint parameterization to replace the homography matrix parameter estimation,achieving planar object tracking.(2)In order to reduce the impact of accumulated errors,this paper uses the siamese neural network to extract features and estimate the homography matrix,and proposes an improved algorithm based on the siamese neural network.The algorithm uses similarity measures to combine the two processes of detection and tracking.The two branches of the siamese neural network extract the features of the template frame and the target frame respectively,and then use the anchor frame mechanism to generate a large number of rectangular regions on the extracted feature maps and sort them.The algorithm divides the planar object tracking into classification and regression tasks at the same time.After classifying the rectangular area,it uses the cosine window and scale penalty to reorder the rectangular area and uses the nonmaximum suppression algorithm to filter.Finally,The filtered results are fine-tuned by regression.After being integrated into the detection process,the target tracking failure phenomenon caused by accumulated errors is improved,and the accuracy of planar object tracking is improved.(3)Thesis uses data samples of five influencing factors including RT(Rotation),SC(Scale Change),OC(Occlusion),PD(Perspective Distortion)and OV(Out of View)on the public data set to quantitatively compare the two proposed algorithms,and compare the proposed algorithm based on the siamese neural network with the field Several representative algorithms have been experimentally compared.The paper presents and analyzes the experimental results of the proposed algorithm in terms of visual effects and quantitative data.Experimental data verifies the effectiveness and accuracy of the algorithm in this paper.