Research On Motorized Target Tracking Sysyem For The Airborne Opto-electronic Platform

Target tracking with complicated background is one of the core technologiesin airborne opto-electronic system, which is used to track motorized targets inairborne environment. This research aims to improve the existing trackingtechnology to improve tracking performance of single target continuously.The existing tracking algorithm of airborne platform is template matching,which can not effectively track the target when encountering complex situations inthe process of target motion, such as temporary occlusion cased by the backgroundobjects, change of rotation or scale, the frequently motion of the camera; and sincethe image matching in spatial domain brings high computational complexity, thereal-time can not be reached. In this study, in order to solve these problems in thetracking process on airborne platforms, the tracking process is divided into fiveparts: the filter estimates, target feature matching, occlusion detection, onlinelearning, and adaption of target size. Filter estimate predicts target location basedon historical information to narrow the range of the search area, so as to increasethe speed; feature matching can accurately detect current position of the targetthrough rapid feature extraction and matching method; occlusion detection is moreaccurate by combining the forecast information, feature matching errorinformation and history matching information; adaption of target size can calculatethe target size in real-time, to adapt to the rotation or scale changes of the targetoccurring in the process of its motion.In this study, we firstly improve the traditional methods based on templatematching in the spatial domain by using the Kalman filter and multi-templatematching method. By using Kalman filter to get the predicted area, and using thefast Fourier transform and integral image to calculate the correlation coefficient,the complexity of computation can be reduced; a matching method that cansimutaniously accomplish template matching with both the background templateand the foreground template is proposed, and combine this method with thematching errors can implement occlusion detection.To further improve the accuracy of prediction, and solve the problem causedby the rotation and scale changes of the target by using template matchingalgorithm, we use condition Kalman filter to estimate, combine the ORB cornermodel with the characteristics of local invariant features and the classification offeatures description to track the target, use the combination of the threshold,matching density and the estimated range to detect occlusion, update target size by using Ball-Snake model and the classification results are used as samples toreal-time learning classifier. Through testing and verification of the algorithms, thetarget tracking task under the requirements of the project is implemented, and itcan continuously track single object in the background of grass and otherenvironments.