Research On Discriminative Object Tracking Algorithm Based On Meta-learning

Single object tracking is one of the critical areas of computer vision,which requires labeling the target state in subsequent video frames given the target information in the first frame.Discriminative object tracking algorithms are favored by researchers in the field of tracking because they make full use of the target background information to improve the discriminability of the model.However,such algorithms still suffer from inadequate exploitation of information from from different feature layers of the backbone,inaccurate target regression,and insufficient model generalization capability.(1)To address the problem of tracking networks that do not make full use of information from different feature layers of the network,this paper uses a discriminative target tracking algorithm based on a multi-classification branch fusion structure to improve the robustness of the model and the discrimination of similar targets to mitigate the target drift phenomenon during tracking,and adopts a meta-learning training method for this branch to enhance the model generalization performance.Secondly,to address the negative impact of inaccurate labeling on the model accuracy in the manually labeled dataset,this paper uses the KL divergence training regression branch based on Monte Carlo sampling to solve the regression problem from a probabilistic perspective.(2)To deal with the problem of insufficient target feature extraction in tracking networks,this paper applies Transformer with a spatial attention mechanism for feature extraction.In addition,this paper applies the energy model to the target tracking regression task,converting the probabilistic regression problem into an optimised energy-based model problem,where the complex integration of the normalisation function is difficult to estimate,and applies an extended noise contrastive estimation to train the energy model to predict the target position more accurately.(3)Focusing on the high complexity of the current Transformer-based target tracking algorithm,this paper explores a novel way of interacting between template frames and search frames by combining a deformable attention-based encoder module with a self-attentive encoder module for feature interaction.The deformable attention-based encoder can accurately track the target location without focusing on all the pixels,which reduces the number of model parameters and effectively improves the model accuracy.