Research On Target Modeling And Updating In Discrimiinative Object Tracking

Because visual tracking theory has more and more applications in surveillance, human-computer interaction, military guidance, medical diagnose and navigation of robots, etc., it has become the key research subject in computer vision. The goal of visual tracking is to make computers able to catch the moving target in the video to prepare for later advanced tasks e.g. video analysis. Based on target’s appearance models, visual target tracking can be categorized into generative and discriminative ones. This paper focuses on target’s appearance modeling and updating as well as classifier design and discriminative feature selection. Firstly, improvements have been made for the updating rule of Haar-like feature, the method of target representation and feature selecting in Multiple Instances Learning (MIL) tracking method. Then, feature selection processing and target model updating are added to the compressive sensing tracking method, which is the modified multi-instance learning tracking method. The main contributions of this paper are summarized in the following:1. A new updating rule of Haar-like feature for MIL trackers to dynamically reflect the appearance changes of the tracked object. In the original MIL tracker, the appearance model is assumed to obey normal distribution and its updating rule consists of a simple linearly weighted sum of the original and the current target distributions in the current frame. However, this updating method is not proved theoretically. In this work, we merge two normal distributions into one in maximum likelihood estimation and get less error than the original one. The average accurate tracing rate of new MIL tracking based on modified updating method is86.78%,21.44%higher than the original MIL tracking. The proposed appearance updating method can enhance the robustness, and be naturally extended to multivariable distributions, useful to track color object.2. Improvements for target representing and feature selection in MIL tracking. Firstly, we adopt Distribution Fields (DF) layer as feature replacing traditional Haar-like one to model the target, exploiting DFs’specificity and landscape smoothness. Secondly, we integrate samples’importance into the weighted-geometric-mean MIL model and derive an online approach to maximize the bag likelihood by AnyBoost gradient framework to select the most discriminative layers. Due to the more discriminative features, our tracker reaches90.14%of average accurate tracing rate in12videos, more robust while needing less features than the traditional Haar-like one and the original DFs one.3. In order to enhance the robustness and speed of compressive sensing tracking in complex scenarios, improvement of tracking algorithm by means of new feature selection and target model updating. Firstly, we select features distinguishing the target from the background, according to the distance between a feature’s positive and negative class conditional probability Gaussian distributions. Secondly, we update target appearance model according to the difference between the current target and the original one, so that the target would not be wrongly updated in case of big occlusion or frequent posture changes. Experiments on ten standard and complex test video sequences demonstrated that for the threes measurements, i.e. center error, success rate and precision plot, our algorithm, with the average correct rate of85.19%,2.87%higher than the original one, archives higher performance than three state-of-the-art methods. The proposed method of feature selection and target model updating, enhances the robustness and the speed.