Research On Single Object Tracking System Based On Discriminative Learning

Object tracking in image sequences is a one of the fundamental research topics in computer vision and artificial intelligence, which can provide the information of scale and position of the object in images. These kind of information is of essential importance in a wild range of practical applications, such as security and surveillance, human-computer interaction, autopilot and control, medical imaging, military and so on. However, suffering from occlusion, rotation, background clustering, fast motion, blurry and other intricate circumstances, the results of current methods cannot be satisfying.In this thesis, we firstly analyze the major frame of present object tracking algorithms, tracking by detection. Inspired by the frame, we start researches from related discriminative learning, which are the core of the tracking frame. So we propose a wellprincipled discriminative dictionary learning(DDL) method, which adaptively builds the relationship between dictionary and class labels by utilizing norm. To be specific, we separatively impose a joint sparsity constraint on the coding vectors of each class to learn the class correspondence and relatedness for the dictionary. This method combines the advantages of the global dictionary learning model and the class-specific dictionary learning model. On this basis, we execute the study of visual tracking. Instead of employing norm on reconstruction term, this technology applies the Huber loss function to make the value of object function growing more slowly as the residue increases. At the same time,the Fisher discriminative term is exploited to make the problem easier to solve. At this point, we achieve the first tracking algorithm and compare it with other methods based on sparse coding. Experimental results on some videos show that our technology has some benefits.Algorithm mentioned above can obtain good performances on some scenes. However, its low efficiency prohibits it from real-time applications. The correlation filter based methods have shown high computational performance, so we commence new technology research and implementation. This time, logistic regression is designed as a filter in Fourier domain, and also applied to kernelized space which can solve nonlinear problems.We utilize Laplace distribute to generate the discriminating labels and features fusion to represent the object's appearances. In order to estimate the scale of the object, we select a suitable one from the candidate size. Then we propose the kernelized logistic regression correlation filter which can achieve real-time tracking and maintain high accuracy. As compared with other state-of-the-art tracking methods, ours performs the best.Finally, we implement the tracking system by c++ language and design our interface with the help of MFC and the open-source library Open CV. The system contains two modes: real-time tracking model on video camera and off-line tracking model on videos. In addition, parameter setting is provided as an extra function in our system, thus conveniently tune the parameters for a specific scene.