The Algorithmic Research On Multi-object Tracking Based On Conditional Random Field

Video tracking is an important task in computer vision,which refers to the continuous inference process of the target state in video sequence.Its task is to generate the target trajectory by locating the target in each frame of video,and provide a complete target area at every moment.Video tracking technology has a very wide range of applications in military and civilian aspects,including unmanned aerial vehicles,precision guidance,air warning,battlefield surveillance,etc.Civil aspects include mobile robot,intelligent video monitoring,intelligent transportation system,human-computer interaction,virtual reality and so on.Video tracking technology refers to the detection of the interested target in the image sequence of video,which is relatively moving with the background,and the corresponding algorithm is adopted to locate the target in each successive image,so as to realize the whole-process tracking and extract parameters.Video tracking is based on video target detection,which is a technology to segment moving targets from video.Target detection is the basis of computer vision technology,which provides conditions for object-oriented video coding,target tracking,motion parameter extraction and other follow-up work.The definition of tracking can be defined as estimating the trajectory of an object in the image plane when it moves around a scene,that is,a tracking system assigns consistent labels to tracking targets of different frames in the same video.Target tracking is not only a challenge,but also an important task in the field of computer vision.With the popularity of cameras,the increasing demand for automatic video analysis has aroused great interest in target tracking algorithms.However,due to the interference brought by camera and background motion,such as occlusion,deformation,background speckle,scaling,etc.,the research of multi-target tracking becomes particularly difficult.At the same time,in the subsequent data association,the traditional optical flow feature or colorfeature is not satisfactory.Therefore,this paper converts the online multi-target tracking problem to the global energy minimization problem by constructing a conditional random field(CRF).At the same time,structural information and depth appearance characteristics are combined to make data association more accurate.In this way,we can closely associate the new video frame detection with the previous target trajectory.In the phase of target detection,the high-precision target selection box is firstly extracted through the Faster RCNN network based on deep learning.In terms of structure,Faster RCNN has integrated feature extraction,candidate box extraction and boundary box regression correction into a network,greatly improving the comprehensive performance and especially the detection speed.In target tracking phase,first for the target character description,mainly using the deep learning method based on similarity measure learning,would expand the positive and negative samples after image projected on the characteristics of public space,and video frames in the same sequence of pedestrians feature vector projection distance in this space,but different characteristic vector projection in this space distance far away.Subsequently,structural information is used for initial data association,and relative position invariance between tracked targets is used to overcome local changes caused by movement,so as to obtain several possible data association schemes.Finally,under the traditional CRF graph model,this paper proposes a model based on the minimum global energy of the trajectory,and constructs a loss function to further constrain the proposed data scheme,so as to achieve target tracking.Finally,in this paper,the target tracking methods based on conditional random field are respectively applied in the static background databases adl-rundle-3,kitti-16,pets09-s2l2,tud-crossing,licence-1 and the dynamic background databases adl-rundle-1,eth-crossing,kitti-19.Experimental results show that the proposed target tracking algorithm based on conditional random field can achieve better tracking results in various tasks.