The Research On Infrared Target Tracking Based On Mean Shift

Infrared imaging technology works by acquiring infrared radiation from outside. It has many advantages such as long operating range, convenience for hiding and ability to work double tides. With the rapid development of infrared imaging technology, it gets a widely use in the military and civilian applications fields, include:infrared precise guidance, video surveillance, reconnaissance and security inspection, search and tracking, and many other aspects. As one of the important techniques, target tracking based on infrared imaging play an important role. Differ from the visible image, the infrared image have the drawbacks of low signal-noise-ration, complex background information, and low intensity difference between target and background, which make it more difficult for robust target tracking in infrared sequences.In this thesis, we do some research on infrared target tracking based on Mean Shift. As one of the commonly used algorithms in dynamic target tracking, Mean Shift tracking algorithm can solve object matching problem between two successive frames effectively without prior knowledge. And it runs both fast and effective. In the Mean Shift tracking algorithm, the model of target is built based on color histogram. Then, Bhattacharyya coefficient is used to measure the similarity between the target model and the candidate model. Finally, the optimal target position can be obtained by iterative calculation.Because of the strong noise, it is hard to distinguish the target from background in infrared images. Although traditional Mean Shift tracking method can be effectively applied in visible images, it makes bad results in infrared tracking. Aimed at this problem, this thesis puts forward the improved Mean Shift tracking algorithm, which can work in infrared tracking situation. In order to enhance the representation capability of target model, we improve the original algorithm from two aspects in this thesis. First, in order to reduce the localization error of object tracking, coefficients based on gray histogram of background pixels around target are computed and incorporated into the computation of target model. Thus, the accuracy of target localization is increased. Second, the texture information, which is free from the interference of illumination and background pixels, is introduced into feature space. We calculate the LBP features and use five texture patterns of them to identify the key points in the target region. After get the texture information, we combine it with gray information to represent the target by calculating the joint histogram, thus making the target representation more robust. Experimental results show that the improved algorithm can achieve robust tracking results in infrared sequences.In the process of infrared target tracking, the scale of target may change constantly. In this situation, the traditional Mean Shift tracking method may fails because of the rigid tracking window. In this thesis, the covariance matrix of tracking window is calculated first. Then we use principal component analysis method to determine the scale and orientation of tracking target. By doing this, we realize the self-adaptive adjustment of tracking window. Experimental results show that our algorithm can accurately obtain the scale change of target in infrared sequences.