鲁棒视觉跟踪中的目标表示与模型更新关键技术研究

项目名称： 鲁棒视觉跟踪中的目标表示与模型更新关键技术研究

项目编号： No.61473309

项目类型： 面上项目

立项/批准年度： 2015

项目学科： 其他

项目作者： 侯志强

作者单位： 西安邮电大学

项目金额： 80万元

中文摘要： 视觉跟踪技术在民用和军事领域有着极其广泛的应用，但在实际复杂环境中，长时间持续稳定跟踪不断发生变化的目标依然是一项极具挑战性的任务，实现这一任务的关键在于提高视觉跟踪算法的鲁棒性。视觉跟踪主要模块包括目标表示、搜索策略和模型更新，搜索策略主要决定跟踪的实时性，而目标表示和模型更新则主要决定跟踪的鲁棒性。为了提高视觉跟踪算法的鲁棒性，本项目将重点研究视觉跟踪中的目标表示和模型更新问题，在传统的视觉跟踪框架内，通过构建一个能够在线存储和处理大量目标先验知识的目标模型先验域，借助基于中层视觉线索和经典模板匹配相结合的生成式模型，以及基于深度学习分类器与经典在线AdaBoost分类器相结合的判别式模型，解决目标表示问题；通过多跟踪器的融合与交互，选择稳定的跟踪结果对目标模型先验域进行更新，从而解决模型更新问题；最终建立一套鲁棒的视觉跟踪算法。本项目的研究将为视觉跟踪技术的发展提供新的方法和思路。

中文关键词： 视觉跟踪；目标表示；模型更新；中层视觉线索；深度学习

英文摘要： Visual tracking is widely used in civil and military areas. However, in a real complex environment, how to cope with the continuously changed target to achieve a long-term robust tracking is still a very challenging task. The key point to resolve the above problem is how to improve the robustness of visual tracking algorithm. A common visual tracking algorithm consists of target representation, search mechanism and model update. The real-time performance of visual tracking relies more on search mechanism, while, the robustness of visual tracking relies more on target representation and model update. In order to improve the robustness of visual tracking algorithm, problems of target representation and model update in visual tracking are focused and researched in this project. Under the framework of the traditional visual tracking, a prior domain of target model which can online store and process plentiful prior information of the target is established. With the prior domain of target model, the generative model which combines middle level visual cues and classical template matching, and the discriminative model which combines deep learning classifier and traditional online AdaBoost classifier are used to solve the target representation problem. By fusing and interacting multiple trackers, robust visual tracking results are selected to update the prior domain of target model and thus to solve the model update problem. Based on these processings, an integrated robust visual tracking algorithm is established. The implementation and expected results of this project will provide new ideas and methods for the developing of visual tracking technology.

英文关键词： visual tracking;target representation;model update;middle level visual cues;deep learning