Object Tracking Algorithm Optimization Based On Recurrent Nerual Network

In recent years,in the field of computer vision,the research of target tracking has become one of the hot topics of discussion.Video single target tracking is an important research topic in computer vision,and has a wide range of applications in video surveillance,robotics,and human-computer interaction.As a key problem in the field of computer vision,the single-target tracking algorithm can draw the trajectory information of a given object based on the tracking information in a continuous image sequence,which is convenient for subsequent behavior analysis and anomaly detection.The advent of the era of big data and the emergence of deep learning methods have provided new opportunities for the research of video target tracking.Therefore,the single target tracking problem has important theoretical significance and research value.In video sequence tracking,the task of target tracking is very complicated,which is mainly reflected in:(1)there are many types of objects in the field of view,and frequent occlusion between different objects;(2)objects will deform when they are moving in the video And size changes;(3)the target will appear to move quickly in the video.These problems bring many challenges to the target tracking algorithm:(1)objects similar to the target in complex backgrounds will interfere with the accuracy of the algorithm;and(2)the tracking of gradual objects in the field of view does not use the visual field information of the video sequence;(3)the advantages of recurrent neural networks in target tracking in deep learning are not fully utilized.Aiming at the above challenges,this paper conducts research on target tracking algorithms.In recent several years,deep learning methods have been successfully applied in the field of object tracking,and have gradually surpassed traditional methods in per-formance.Recurrent neural networks belong to a large category of neural networks and are mainly used for processing information in temporal domain.Nowadays,most of the recurrent neural networks have outstanding results in speech process-ing,which are suitable for data sequences with continuous information.Recurrent neural networks have memory,parameter sharing,and Turing completeness.There-fore,learning the nonlinear characteristics of the sequence has certain advantages.The video sequence of the target tracking has continuous time sequence.At present,there have been related researches in the sequence image processing appli-cation of target tracking,but its characteristics have not been fully utilized.This topic is based on the development of dynamic recurrent neural networks,and studies how to improve the accuracy and processing speed of target tracking algorithms.The main research work and innovations are as follows:(1)A Bidirectional Dynamic Tracking Algorithm based on Recurrent Reural Network is proposed.This algorithm use a dynamic recurrent neural network,a long short-term memory network,as the backbone network,and trains a discriminative tracker to solve the problem of how to make full use of temporal-domain information in complex backgrounds.Through the feature fusion of inter-frame detection,the YOLO detector is used as the preliminary tracking module,which improves the accuracy of target tracking in detecting any target in a single frame.A strategy for bidirectional positioning of targets using a long short-term memory network is proposed to make use of the spatiotemporal continuous information around the target in video information.The experimental results show that compared with the existing algorithms,the proposed algorithm makes full use of the time domain information in the video,and improves the performance of the single target tracking algorithm on the public data set.(2)A Tracking Algorithm based on the Dynamic Siamese Recurrent Network is proposed.This algorithm is mainly based on the Siamese network architecture and uses the ideas of template matching and similarity measurement to calculate the corresponding best matching point by using the matching response score.The algorithm uses convolutional neural networks to extract hierarchical features and performs deep fusion with deep features in recurrent neural networks.At the same time,the recurrent neural network is used as the positioning backbone network to make full use of the spatiotemporal continuous information around the target in video information.The experimental results show that compared with the existing algorithms,the proposed algorithm solves the problem of how to reduce similar background interference to a certain extent,and improves the performance and robustness of the single target tracking algorithm.