| With the continuous development of deep learning technology and artificial intelligence,planar object tracking has become a research hotspot in the field of computer vision,and has extensive application prospects in robot vision and face recognition.In recent years,convolutional neural networks have become one of the core technologies in the field of computer vision.More and more scholars have applied convolutional neural networks to the research of planar object tracking.The existing methods have achieved better research results,but there are still some problems such as high time complexity of tracking and inaccurate tracking.In order to improve the speed and accuracy of planar object tracking,this thesis mainly does the following research work:(1)A single target tracking algorithm based on dense network is given to address the problem that the speed of target tracking needs to be improved.First of all,the fully convolutional dense network is used as the backbone network for feature extraction to generate a dense feature map.Secondly,the target classification based on discriminant model is proposed on the generated feature map to do a rough localization of the target.Then,the overlap maximization method is used to estimate the target boundary box at the rough localization,and the target tracking can be achieved.Finally,the comparison experiments of tracking accuracy and tracking speed are designed on four kinds of benchmark data sets,and the experimental results show that the algorithm is effective.(2)The coarse to fine planar object tracking algorithm is given for the problem that the accuracy of planar object tracking needs to be improved.First of all,the target bounding box is estimated according to the single target tracking algorithm using overlap maximization to achieve coarse localization of the target.Secondly,SIFT is used to locate the key points in the image,and the corresponding feature descriptors of the key points are extracted from the feature map generated based on the cascaded thick-to-thin dense network.The feature descriptors are matched according to the nearest neighbor algorithm to obtain the set of paired key points.Then,the roughly positioned bounding box is used as the prior box to filter the matching key points from coarse to fine,and the key points in the prior box are retained to obtain a more effective set of key points matching.The optimal homography matrix parameters of the key points are estimated by RANSAC algorithm.Finally,the final prediction box of the target is obtained by calculating the ground truth through homography matrix.In this thesis,experiments are designed on POT benchmark dataset to verify the effectiveness of the coarse-to-fine planar object tracking algorithm. |
