| Object tracking is one of the hot and difficult topics in the field of computer vision.Although researchers have come up with a number of tracking algorithms and achieved fruitful results in recent decades,few tracking algorithms achieve the desired performance due to various factors.The object tracking is also an indispensable intermediate step for many high-level computer vision tasks such as video surveillance,video retrieval,human-computer interaction,traffic monitoring,car navigation and augmented reality.Therefore,it is important to research on this topic with a very important theory meaning and practical significance.Aiming at the problems existing in the task of object tracking,this paper proposes a training method based on convolution neural network with center contrastive loss function and a tracking framework with priority of small motions.The main contents and innovations are as follows:(1)In order to solve the problem of weak ability when employing handcrafted features,a convolutional neural network is introduced to extract features.In the selection of convolution neural network model,attention is paid to the semantics of low-level features and high-level features,meeting the needs of the object-tracking task to track a specific target.(2)Design the center contrastive loss function.By calculating the contrastive loss to solve the problem that there is no definite category for any sample in the object-tracking task.In addition,different types of samples are packed to feed into the network,which makes it possible to calculate multiple types of inputs in a single network and simplify the input size and data preprocessing.(3)The motion of the bounding box of the object is modeled,and a tracking framework based on the similarity of images and the priority of small space motion is proposed.In this paper,a full simulation experiment has been carried out.The quantitative and qualitative evaluation of the algorithm also proved the effectiveness of the algorithm. |
PDF Full Text Request