Study On Robust Tracking And Recognition Of Infrared Target

Infrared imaging and the involved image processing technique is one of the key techniques in the application field of missile guidance, automatic search and tracking, infrared alarm and early warning etc. Moreover, it is also an active topic of computer vision and pattern recognition field. As an important research theme of this topic, infrared imaging target tracking and recognition technique receive much interest and attention. With the development of scientific technique, the researchers have proposed a lot of new means and methods of infrared imaging target tracking and recognition which have been widely applied in many real weapon systems. From the state of the art of the corresponding technique published in domestic and foreign, we investigate deeply how to achieve robust infrared target tracking and recognition performance in this dissertation.According to the information difference caused by the varied infrared imaging target size, we divide the tracking issue into two parts, one for small target tracking and the other for area target tracking. For small target, since the target information utilized is only the observed location and gray information, we take H infinity filter as the prediction filter for motion analysis and tracking to adapt the uncertainty of noise source and exotic disturbance in multi-application environment. For area target, based on the discussion of the shortage of single kernel and cascaded kernel mean shift tracking algorithm, we compensate and optimize the performance of the tracking algorithm with combining the online analysis and evaluation judgement of the tracking performance. And the case of target occlusion and size change is also analyzed and handled during the area target tracking process. In the part of infrared target recognition, we develop an infrared target extraction method based on multi-information incorporated kernel density estimation to separate target and background completely. The corresponding feature description vector of the target is inputted into a regularized Adaboost based infrared target recognition framework to obtain robust target recognition performance. The aim of the regularization is to avoid classifier overfitting which incurred by highly noisy infrared image feature. Because of the inseparability of target tracking and recognition in real application, we present an integrated design scheme of tracking and recognition in the application of infrared homing antiaircraft missile which ensures the integrality of the entire dissertation with respect to the current project research.Specifically, the main contributions of this dissertation are as follows:(1) An image pre-processing method with multi-scale Top hat filter is presented to reduce the influence of infrared noise and clutter, which makes the obtained small target observed value more trustful and accurate. H infinity filter is introduced as a prediction filter for infrared target tracking to predict and track the small target, which strengthens the adaptive capacity of the tracker when uncertainty of noise source and exotic disturbance exist in multi-application environment.(2) We design a scheme to construct target model with a cascaded kernel incorporating the spatial relation and gray value information of pixels with which the tracking performance is somewhat improved. Online analysis of cascaded kernel based infrared target tracking performance is achieved by a kernel based tracking performance evaluation metric which unifies the tracking and performance metric within a kernel based framework. We develop a new eigenvalue-based similarity measure to implement position correction in the frame associated with a poor tracking performance. This similarity measure contains not only the gray information but also the spatial relation information of the pixels. In order to compare and assess the performance of the proposed performance feedback compensation kernel-based tracker and the mean shift tracker quantitatively, ROC curve analysis method is introduced for evaluating tracking performance. This provides a new way to analyze the performance of a kernel based-tracking algorithm.(3) H infinity filter is taken as the prediction filter to make up the deficiency of the proposed performance feedback compensation kernel-based area target tracking algorithm in the event of total occlusion, which makes the algorithm robust to total occlusion of the target. Meanwhile, we set tracking window size according to a function of the zero moment of the pixel gray values in the target region which makes the tracking algorithm robust even though target size varies.(4) We design a probability density kernel, which incorporates multi-information, including gray value, spatial relation and local standard deviation information, and it is used for infrared target extraction. A regularized Adaboost algorithm is introduced as classifier with the NaiveBayes as the weaker learner for infrared target recognition with which the target recognition reliability is improved when the data is highly noisy.(5) Combining the reality of current research project and taking homing antiaircraft missile as an example, three basic target recognition criterions: position criterion, area criterion and gray criterion are reinforced for fully utilizing the varied position, area and gray information of the recognized target, which improve recognition performance of the infrared target classifier. With analyzing the basic function of antiaircraft missile weapon system, we present a theory idea of the tracking and recognition algorithms'cooperation mechanism of infrared image information processing part of homing antiaircraft missile.