Research On Object Tracking Algorithms Based On Improved TLD

Intelligent video surveillance is a new application direction of computer domain and focus,it uses the method of image processing and computer vision to localize,identify and track,and then,it can analyze the target behavior in the foundation.It is certain that the moving object tracking technology will have broad application prospects as the core of intelligent monitoring technology.For object tracking,the key issue is to ensure that it can catch the target in the current frame and not miss the object in the next frame.Target tracking algorithm based on testing has far exceeded the single tracking algorithm,so this paper will study on object detection and object tracking,the key problem is to find better ways to describe object effectively and detect or track object in real time.For the object detection system,in the training stage,it uses more scale sliding window method to scan and search these images firstly.Calculate the Histogram of Oriented Gradient(HOG)feature and the Semantic Local Binary Pattern(SLBP)feature on each sliding window respectively.Then,it uses Principal Component Analysis(PCA)to reduce the dimensionality of HOG feature to specific dimensionality.Combine HOG feature and SLBP feature to joint features,then,the linear support vector machine(SVM)would train classifier.Finally,context information would optimize the detection bounding box.Experiments show that the improved method has advantages of the high accuracy,the strong anti-jamming ability,the minor calculation and so on.For the object tracking system,this paper uses the Tracking-Learning-Detection(TLD)tracking framework.TLD algorithm could track object better for a long time in partial occlusion,deformation,and uniform illumination,but it would track mistakenly or fail to track in the case of uneven illumination,serious occlusion or the fuzzy tracking object and so on.For these disadvantages,it will increase the judgment module in the detection module.When the object has a better texture,SLBP classifier will replace the nearest neighbor classifier and convert the image to the SLBP feature vector,and the vector could classify these samples.Experiments show that the TLD-SLBP algorithm obtains a higher success rate than other algorithms and good tracking ability in multiple video sequences.