Object Tracking Based On Graph Representation And Learning

Nowadays,artificial intelligence technology has gradually become the core of the development of science and technology.With the advent of artificial intelligence era,more and more cities are building smart cities which based on artificial intelligence technology.How to design intelligent video analysis technology to deal with massive surveillance video data quickly is the key step in the process of smart city construction,and object tracking is the basic algorithm in intelligent video analysis technology.Therefore,futher research of object tracking is not only the need of the development of computer vision,but also effectively accelerating the construction of smart city.With the development of traditional machine learning method,especially the breakthrough of deep learning method in recent years,object tracking has made rapid progress,and many efficient tracking algorithms have been proposed.The basic process of object tracking generally include:input a new video frame;extract and process the target features;construct the tracking model;locate and tracking the target.Extract and process the target features and construct the tracking model are the key steps of object tracking.Starting from the extract and process the target features,this thesis studies the existing problems in object tracking task,and proposes an effective expression method for tracking target,which can significantly reduce the background noise in object bounding box and improve the robustness of tracking results.The specific methods are as follow:first,we divide the object bounding box into several non-overlapping image patches.Then,each object image patch is weighted by an effective weight computing model,that is,the background image patches are given small weights,and the target image patches are given large weights,which can effectively reduce the influence of background noise.Finally,we combine the image patch weights with the image feature to obtain the weighted feature descriptor of the object image patches,which is taken as the representation of the tracking target so as to realize the object tracking.Among them,the weight computing model is the key of the weighted feature descriptor.This thesis analyzes the tracking target,background noise and so on,and proposes the following three effective weight computing models:First,observing the process of object tracking,we find that the spatial structure of object is the key information to locate the tracking target.The traditional graph representation model based on the nearest neighbor graph,which only consider the local spatial information of the tracking target,and ignore the global information.This will make the representation of tracking target lack unity.if the appearance of the tracking target is complex,the traditional graph representation model cannot get the effective feature representation and directly affect the tracking results.In this thesis,we propose Laplace regularization random walk ranking model by adding Laplace regularization constraints to the random walk with restart,which makes the local and global spatial information be fused together in the process of solving the weights,which make the weights distribution conform to the target spatial structure.Finally,the weighted feature descriptor of the object image patches is used as a representation of the tracking target and is combined with the structured support vector machine tracking model to achieve the final object tracking algorithm.In this thesis,object tracking algorithm is implemented on two large-scale benchmark OT100[32]and TCL128[59],and the tracking results are compared with many tracking algorithms proposed in recent years.The experiment results show that the tracking algorithm of this thesis achieve the best performance on two datasets,especially when there are fast motion,deformation,rotation,illumination changes,motion blur and other challenges which directly lead to the appearance change of tracking target,the tracking algorithm still maintain robust tracking results.Second,during the tracking process,most of the tracking targets and the background have significant difference.Therefore,using background information to highlight the tracking target will not only effectively reduce the impact of background noise,but also highlighting the tracking target,In this thesis,the conditional random field model is used to jointly optimize the background information,foreground information and the spatial structure of the tracking target to obtain a more discriminative weights distribution.Finally,the weighted feature descriptor of the object image patches is used as the representation of the tracking target and use the structured SVM algorithm as the basic tracking framework to achieve the final object tracking algorithm,and implemented on two large-scale benchmark OTB100[32],TCL128[59].The tracking algorithm compare with many recent tracking algorithms.The experiment results show that the tracking algorithm has the best comprehensive performance on two datasets,which including the tracking speed.In particular,the tracking algorithm can achieve accurate tracking results when tracking target is affected by background clutter,partial occlusion,and low resolution.Third,the constructed graph structure has a good expressive ability or not will directly affect the results of the graph representation model.In addition,the original feature information of the image patches is only used in the process of graph construction,and ignore the association between the feature of image patches.This thesis proposes a graph representation and semi-supervised learning model to dynamically optimize the association between image patches,and integrates the feature information of the image patches into the process of weight computing to make the weights distribution more in line with the features of tracking target.Furthermore,we proposes a fast and efficient iterative optimization algorithm to solve the graph representation and semi-supervised learning model.During the tracking process,object scale estimation and object re-detection technology are add into the tracking framework.The tracking algorithm compare with many recent tracking algorithms on three large-scale benchmark OTB 100[32],TCL128[59]and VOT2015[58].On OTB100[321 and TCL1281591,the tracking algorithm of this thesis has obtained the best comprehensive tracking results,and has excellent performance in many challenging attributes.On the VOT2015[58]dataset,the algorithm of also obtained comparable results.