目标跟踪中目标模型更新问题的半监督学习算法研究

本文针对长期稳定的目标跟踪中的目标形变、尺度缩放、旋转等问题, 提出一种步步为营的反馈式学习方法, 该方法通过正、负约束实现对于目标模型和分类器的判别能力和容错能力提高的同时, 使更新带来的误差尽量小, 并证明了其收敛性. 通过实验表明, 对于同一种跟踪算法使用本文提出的目标更新方法进行更新学习的比不更新学习的跟踪效果要稳定得多, 对于目标的尺度变化、形变、旋转、视角变化、模糊等都有较好的适应性, 并通过与现有的较流行的方法进行比较, 本文方法鲁棒性较好, 有很高的研究和应用价值.

Research on semi-sup ervising learning algorithm for target mo del up dating in target tracking

Target detection and tracking technique is one of the hot subjects in image processing and computer vision fields, which has significant research value not only in military areas such as imaging guidance and military target tracking, but also for civil use such as security and monitoring and the intelligent man-machine interaction. In this paper, for target deformation, scale changing, rotation, and other issues in the long-term stable target tracking, a bootstrapping feedback learning algorithm is proposed, which may improve the target model and the classifier discriminating capacity as well as the fault tolerance ability; and it also makes fewer errors during the updating, and then the proof of convergence of the algorithm is given. Experimental results show that among the same tracking algorithms, utilization of the learning method to update the target model and classifier is more stable and more adaptable than unusing it in the processes of target scale changing, deformation, rotation, perspective changing and fuzzy. And compared with the existing conventional method, this method has a better robustness, and a high value in practical application and research.