Research On Target Tracking Based On Siamese Networks

Target tracking is an important branch of computer vision,which has been paid attention by researchers all over the world.Video target tracking algorithms can be widely used in various real-world scenarios,such as: virtual reality,unmanned driving,intelligent transportation,and human-computer interaction.The task of video target tracking is to give a specified target in the initial frame of a video sequence and predict its size and position in subsequent frames.However,at the current stage of target tracking,many difficulties and challenges are encountered,such as changes in lighting,motion blur,occlusion,and rotation.These difficulties and challenges will affect the accuracy of the target tracking algorithm and may cause drifting in the tracking target.Therefore,based on the above relevant knowledge,the target tracking continues to be studied in depth.The general research direction of target tracking algorithms is divided into two categories,one is the traditional method,and the other is the method based on deep learning.Traditional target tracking methods mostly use correlation filtering,extracting image blocks at the target position given in the first frame,and training to get the correlation filter.As deep learning is more and more widely used in the field of computer vision,the algorithm of deep learning is similar to the process of human brain processing algorithms,because the brain and deep learning models involve a large number of computing units.The target tracking algorithm based on deep learning is gradually developed,and various neural network algorithms are added to it,which can greatly improve the robustness of the target tracking algorithm and accurately track the target in the video.This article aims at three problems in the process of deep learning training(1)deep learning training takes more time.(2)The parameters of deep learning have different attributes.(3)Deep learning requires a lot of data in the training process.A deep learning parameter evaluation method based on evidence reasoning is proposed and constructed.The evaluation framework is a hierarchical structure,which is divided into three layers.The evidence inference algorithm is added to it.Before each parameter modification,the parameters are entered into the evidence inference parameter evaluation algorithm to determine whether the set of parameters can improve the robustness of the algorithm,which can greatly reduce the time wasted training algorithm.In Chapter 4,the target tracking algorithm based on the attention mechanism and the siamese network is proposed.It is improved on the basis of the SiamFC algorithm.The added convolutional attention module divides attention into two dimensions,spatial attention and channel attention.Simultaneously input two pictures into the network and compare their similarity.If the similarity is high,the score is relatively high.The position with a high similarity score is used as the target position for prediction in the next frame.The algorithm is improved by 0.6 percentage points based on SiamFC.