CAREER: Visual Tracking with Online and Prior Learning

职业：通过在线和先前学习进行视觉跟踪

• 批准号: 1149783
• 负责人: Ming-Hsuan Yang
• 金额: $ 47.38万
• 依托单位: University of California - Merced
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 2012
• 资助国家: 美国
• 起止时间: 2012-01-01 至 2022-09-30
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1149783&HistoricalAwards=false
• 关键词: CAREER; Visual; Tracking; Online; Prior

This project develops efficient and effective algorithms to handle challenging problems in visual tracking such as drift, heavy occlusion, and failure recovery. The research team is developing an integrated framework in which object detection, tracking and recognition are addressed simultaneously. Within this framework, the prior knowledge is learned from a large set of images pertaining to object classes of interest. Such knowledge serves as long-term memory for the proposed appearance models which are then adapted to unseen new object instances. In addition, a top-down saliency model for each object class of interest is developed in order to handle heavy occlusion and failure recovery. The project has four major components: developing algorithms for learning visual prior and transferring knowledge for online appearance model, designing tracking algorithms that handle draft with the proposed appearance model, modeling top-down saliency maps to handle full occlusion and tracking failure, evaluating state-of-the-art algorithms with a large benchmark dataset. This project provides a building block for robust object tracking, which can be applied to motion analysis, surveillance, and multi-object tracking. The developed top-down saliency map provides a flexible way to represent objects, which can be extended to object detection and segmentation. The proposed tracking library and benchmark data set provide a platform for evaluation of advances in object tracking. This research is integrated with education and outreach by courses and activities aimed at attracting students to this field and encouraging interdisciplinary collaborations.

该项目开发了高效有效的算法来处理视觉跟踪中的挑战性问题，如漂移、严重遮挡和故障恢复。研究小组正在开发一种综合框架，其中同时处理目标检测、跟踪和识别。在该框架内，从与感兴趣对象类有关的大量图像中学习先验知识。这样的知识用作所建议的外观模型的长期记忆，然后将其适应于看不见的新对象实例。此外，为了处理严重遮挡和故障恢复，为每个感兴趣的对象类开发了自上而下的显著模型。该项目包括四个主要部分：开发用于在线外观模型的视觉先验学习和知识传递的算法；设计使用所提出的外观模型处理草稿的跟踪算法；对自上而下的显著图进行建模以处理完全遮挡和跟踪失败；使用大型基准数据集评估最新的算法。该项目为健壮的目标跟踪提供了一个构建块，可以应用于运动分析、监视和多目标跟踪。开发的自上而下的显著图提供了一种灵活的对象表示方式，可以扩展到对象检测和分割。拟议的跟踪库和基准数据集为评价目标跟踪方面的进展提供了一个平台。这项研究通过旨在吸引学生进入这一领域并鼓励跨学科合作的课程和活动，与教育和外展相结合。