The Research On Siamese Net And Meta Learning Based Visual Object Tracking Algorithm

In recent years,the development of artificial intelligence is extremely rapid.As one of the most important methods of intelligent video image sequences analysis and processing,visual object tracking has gradually become a hot research area in the field of computer vision,and attracts extensive attention.Visual object tracking has broad development and application prospects in many scenarios such as traffic supervision,unmanned driving,intelligent navigation,human-computer interaction and even military field.Its related research results have been widely used in the construction of various intelligent video image analysis systems,which is the important driving forces to promote the development of artificial intelligence and has great research significance.Visual object tracking task can be roughly described as tracking and locating the interested object in a video or image sequence,and calculating the coordinates of the target object in each image.Because of many disturbances in the natural environment,tracking algorithms are faced with many difficulties and challenges,such as deformation,rotation,sheltering,scale transformation,illumination change,motion blurring,interference from similar objects,etc.In this thesis,a visual object tracking model based on Siamese-structured convolutional neural network is proposed,which improved the tracking accuracy and robustness in the face of various interference challenges in natural scenes by using the meta-learning gradient descent based optimization method.Combining with the text tracking task which has less intersection with visual object tracking,we make corresponding improvements to its difficulties and realize the application of tracking model in text scene.In summary,this thesis includes following contributions:1.Based on the current mainstream Siamese network tracking model,a new object tracking algorithm based on meta-learning gradient descent is proposed.Model made full use of the monitoring information of target coordinates in template frame by updating the convolutional kernel's parameters in regression branch,so that the coordinates calculated by tracking model can fit the ground truth of target more accurately.2.The Model-Agnostic Meta-Learning training method is adopted to update the parameters.Through this method,the model can obtain the optimal parameters that adapt to the current video with only one or a few steps of gradient descent in the first frame image when faced with new tracking video,so that the algorithm can implement good tracking results in current video.3.According to the characteristics of text video scenes and the particularity of text objects,customized improvements such as feature enhancement,mask attention,online updating is made to realize the application of visual object tracking algorithm in text scenes,which has some impetus to both research of object tracking and text tracking.