Research On Visual Object Tracking Algorithm Based On Deep Learning

Visual target tracking is the research focus in the field of computer vision.Its development is of great significance and has been widely applied to various fields,such as aerospace,defense construction,medical and health.Although significant progress has been made in tracking algorithm research over the past few decades,achieving continuous tracking is still a huge challenge due to the complex and variable tracking scenarios.In recent years,deep learning has been introduced into target tracking research and has made breakthrough progress.Deep learning is a multi-layer neural network that simulates the process of human brain learning and analysis.By combining low-level features to form a more abstract high-level representation,this distributed architecture can better learn the essential features of objects.This thesis uses a deep neural network model with powerful feature extraction ability to describe the appearance of the target.In view of the fast motion,background clutter,rotation,illumination variation and other issues affecting the target continuous tracking,from the perspective of how to use the features from multiple convolutional layers to construct an effective appearance model,some new tracking methods are proposed,the main research ideas and innovations are as follows:(1)A target tracking method based on spatial context pyramid is proposed.This method develops spatial context information to better learn the relationship between target and background,and designs an effective space window to suppress the background information while retaining the target information,which enhances the discriminating ability of the tracker.The joint use of the multi-level space window constructs a context pyramid appearance model,and each level of the pyramid contains different levels of context information to accommodate different challenge factors.In order to verify the effectiveness of the algorithm,the versions of depth features and traditional features are implemented to compare with representative methods.The experimental results show that the proposed algorithm has better tracking accuracy for fast motion and background clutter problems.(2)A target tracking method based on historical retrospect is proposed.The method firstly studies the influence of features extracted from different convolutional neural network layers on visual tracking problems,and then designs a historical retrospect verification mechanism.This mechanism judges the reliability of tracking prediction and updated model by bidirectionally locating the target and calculating deviations.At the same time,this deviation is also used as a criterion for convolutional layer selection,and if necessary,features will be re-selected to reposition the target.A large number of experimental results show that the proposed algorithm has better tracking accuracy for background clutter,rotation,illumination variation and other issues.(3)A deep feature tracking method based on interactive multiple model is proposed.The method applies the correlation filter to the output of multiple convolutional layers to construct an observation model,and learns a corresponding online motion model for each observation model.These observation models and motion models form multiple model systems.The interactive multiple model is then applied to the visual tracking to adaptively adjust the weight of each model system to achieve robust tracking.In order to verify the effectiveness of the proposed method,experiments and evaluations are carried out on two public target tracking datasets.The results show that the proposed algorithm has better tracking accuracy for background clutter,rotation and illumination variation.