Research On Moving Target Tracking Algorithm In Video Sequences

Visual targets tracking is the key technology of computer vision,it is also the research foundation of artificial intelligence,and play an important role with a broad development prospects in the intelligent monitoring,medical diagnosis,human-computer interaction and other fields.At present,these problems of visual tracking have been studied extensively and deeply,and have achieved great success.However,with the development of science and technology and the popularization of image equipment,the higher performance target tracking technology is needed to deal with the task of tracking in complex scene,so there are still many problems to be solved,such as deformation,illumination Variation,scale variation and fast motion.The design of stable and efficient target tracking methods is still full of challenges.The purpose of this thesis is to study correlation filter-based tracking algorithm,and improve the tracking accuracy to a great extent.An adaptive scale target tracking algorithm based on SVM is presented and implemented in this thesis,some experiments are carried out to show the efficiency of the proposed algorithm by using a set of video sequences.Main work of this thesis includs the following aspects:1.This thesis analyzed the research status of target tracking algorithm from three aspects:description of appearance model,motion model and model updating,and summarized the technical difficulties and the disadvantages of the existing tracking methods.2.In this thesis the principle of correlation filter and the method of kernel correlation filter tracking are for intensive study,the filter method is used to construct a scale filter to estimate the target scale,and the tracking task is decomposed into translation estimation and scale estimation.At the same time,considering the target's scale is assumed to not change much between consecutive frames,and take advantage of a priori probability to adjust the response of scale filter,this strategy will produces more stable detections.3.In order to suppress the influence of the window function caused by the scale variations,this thesis proposed a method using the adaptive window function to filter the target signal.When the scale of target changes,the window function of the fixed bandwidth parameters will alter the signal to process and affect the learning process.Using the Gauss function of the adaptive bandwidth parameter as the window function to eliminate the influence of the target signal due to scale change,thus improving the stability of the tracking model..4.A robust tracking algorithm must have the ability to track for long time in the case of tracking failure.In order to handle occlusion problem,a re-detection module was introduced to correlation filter-based tracking framework,by training an online detector based on SVM to re-detect,when the detection of the target occlusion occurs,the re-detection mechanism is performed,the optimal candidate result is selected as the result,and adaptive adjust the learning rate of model.When the target appeared again after occlusion,the tracker could still track the target effectively.At last,a conclusion was made on the research of this thesis,the prospect which laid the foundation for the following research work was made on the further plan.