Single target tracking is one of the most challenging application scenarios in the field of computer vision. To solve the problems of occlusion

object deformation

and motion blur during tracking

a training method was proposed to train single target tracking network based on generated diverse positive instances

and the problem of scarcity of various training samples was also mitigated. Specifically

during the offline stage

a variational autoencoder (VAE) was employed to encode original samples into latent space. Then

the hard positive data was generated via sampling variables in latent space to improve the diversity of the training data

and a training dataset was constructed by combining the generated data and the original samples. Besides

for the target template and the negative and positive samples of the training sequences

a probability triple loss function was utilized to train the tracking network. The relation between the positive and negative samples was investigated to improve the discriminative power of the tracking network. During the test stage

the pretrained Siamese neural network (SNN) was used to track the target

and the position of the target at the moment could be determined by correlating the target template and the search area. The experiment shows that the proposed algorithm improved the robustness and accuracy of SNN tracking in the cases of interference of similar objects and deformation

fast motion

rotation

motion blur

and occlusion of target during movement

and achieved real-time tracking performance.