Research On Visual Moving Object Detection And Tracking

Visual moving object detection and tracking is a heavy study area in computer vision, becauseit is a important part of many computer vision application，such as video surveillance,3D redi-stribution, and so on. Compared with30years ago, visual moving object detection and trackingtechnologies have a great improve on the aspect of compute time, accuracy and robustness with theeffort of many scholars. However, due to the complexity of the real world and the limits of humanknowledge, the moving object detection and tracking technologies still have many difficulties. Inorder to solve these problems, I study some algorithm in this thesis. My major works are as fol-lows:1. Study some classical moving object detection algorithms, e.g. Gauss, MOG, KDE, Code-book, ViBe and LBP, and then discuss the advantages and disadvantages of each of these algo-rithms from both theory and experiment. Secondly, analyses several multi-object tracking algo-rithms, e.g. Linear Program, k-short paths and GMCP, which recently proposed from the combina-tion optimization theory, and then prove the feasibility of these algorithms in experiment.2. A moving object detection algorithm using the Scale Invariant Local Ternary Pattern(SILTP) and VIBE (VIsual Background Extractor). The first frame of the video is used for con-struct the background model, and then the blind and selective update mechanisms are consideredinto the model update by the way of random alternative, and last the similarity of the SILTP is usedfor discriminate the background and foreground pixel. Experiment shows that the proposed algo-rithm is very effect to handle many screens, and the compute time and accuracy of the algorithmare both better than others.3. A global multi-object tracking algorithm using hierarchical data association following thetracking by detection framework. We first in the whole video use an object detector to obtain thedetection response，and then utilize the Generalized Minimum Clique Graphs (GMCG) to solve thedata association problem on detection response in video clip, this step can obtain tracklets. At thelast we obtain the object track by solving the association problem on tracklets in whole video usinga hierarchical method. Experiments on the public datasets show the proposed method can solvedata association and handle occlusion effectively.