Research On Object Tracking Algorithm Based On Neural Network

Object tracking is one of the most important technologies in computer vision,its purpose is to monitor the object of the next frame in video sequences,according to the initialization of given an object frame.The task is to generate the trajectory and provide complete area of object through object position of each frame in the videos.Object tracking technology combines image processing techniques,image detection techniques,image recognition techniques and automatic control techniques,and is widely used in intelligent monitoring system,intelligent traffic,automatic driving,battlefield surveillance,medical image and human computer interaction.The traditional object tracking algorithm uses correlation filter to extract and match features,and its tracker has some limitations,which can only track the object in a specific scene.With the development of neural network,it be adopted to extract feature.although the object tracking algorithm have made great progress through the neural network for feature extraction.Actually,object tracking is also a very challenging task due to the influence of many factors,such as illumination,background clutter,scale variation,fast motion,deformation,occlusion,and rotation etc.in practical application.In view of the fact that the accuracy and robustness of the object tracking algorithm are not ideal,this thesis proposes an object tracking algorithm based on the deep siamese network.The algorithm on the basis of the overall framework based on siamese network.Firstly,residual neural network resnet50 is used to extract feature of template frame and the object frame,on the basis of the features fusion in the convolution layer,using feature pyramid networks FPN to mix shallow and deep feature between convolutional blocks to improve accuracy and precision in the condition of deformation and the rotation.Secondly,the observation model of object tracking has been designed,through adding the object mask module to the template frame and the object frame of pixel matching to draw an object outline,converting the apparent features to the pixel features,the increasing of the object tracking algorithm application range and improving robustness of object tracking algorithm.Finally,the minimum rectangle fitting is used to express the object appearance,which improves the occupancy ratio between positive samples and negative samples in the prediction box,and further improving the object tracking accuracy.In order to prove the effectiveness of the proposed algorithm,the performance of the proposed algorithm was tested and compared on the test data sets OTB2015,VOT2018 and DAVIS2016,which are recognized as different benchmarks.Experimental results show that the algorithm in this thesis achieves the best performance on multiple data sets under different standards.The average precision is improved by 1%,and the robustness is also improved.Compared with the existing object tracking algorithms based on siamese networks,the proposed algorithm can effectively improve the location of object in the condition of deformation and rotation,and achieve the effect of improving the accuracy and robustness of the proposed algorithm.