| Object detection and target tracking are one of the hot topics in computer vision and video processing.With the rapid development of artificial intelligence and big data,in many areas such as video security,autonomous driving,virtual reality,image understanding,robot control,visual control require their research and development.In real life,due to the need to ensure the accu rate detection and tracking required in a variety of application scenarios,and the need for real-time detection and tracking,this brings great challenges to this research field.Target tracking is to continuously find the target that needs to be tracked in the video sequence.The tracking algorithm as a whole also starts from traditional feature-based extraction and machine learning to deep learning-based neural network tracking.In the field of target detection and tracking in recent years,end-to-end convolutional neural networks based on deep learning have developed rapidly,especially neural network target trackers combined with related filtering methods that originated in the field of signal processing.As far as computer vision is concerned,target tracking based on convolutional neural networks does not need to define features by themselves.Deep learning networks can have powerful description capabilities for features and can learn features by themselves.Because of this,the end-to-end deep learning framework network tracking structure is widely used.Based on correlation filtering methods and attention mechanism,we have conducted in-depth research on target detection and tracking.The goal is to improve tracking accuracy,enhance tracking robustness,and at the same time ensure tracking speed.We use the deep and shallow features of the deep convolutional neural network,and make full use of the feature information and semantic information of different layers.A method based on siamese networks combining channels and spatial attention models is proposed to make full use of channels and spatial dimensions to improve the accuracy of the tracking network.The research work of this paper is as follows.We use Alex Net proposed by Alex et al.as a benchmark feature extractor for tracking twin networks,introduce the RPN network idea in target detection,and perform end-to-end target tracker training from classification and regression two branches to visually tracking issues are viewed as a cross-correlated issue.For the problem of imbalance between classification loss and regression loss in the training process,the loss function is designed to speed up the training convergence and learning speed of the network model,and further improve the performance of the target tracking network model.Discuss the feasibility of using deep neural network architecture in the field of target tracking,and use spatial sensing sampling strategy and multi-layer information fusion to solve the problems of deep network for tracking network. |
