基于多任务学习的跨视角智能视频分析方法研究

Intelligent video analysis plays an increasingly important role in everyday life with all kinds of real applications. Cross view is one of the most challenging problems in video analysis, which limits the performance of algorithms and applications in multi-scene data analysis. In this proposal, we will focus on cross view video analysis and try to improve effectiveness and efficiency of algorithms. With the summarization of previous approaches, we believe it is a good solution for cross view video analysis to connect respective vision tasks in different views by all kinds of cues to formulate a unified multi-task learning framework. In this way, we can improve the communication of different views and mining the complementary information. The contents of the research in this proposal include: an deep learning based attribute extraction and hierarchical feature construction; multi-task learning based on different cues; fusion of cues to improve the performance of video analysis; a unified framework based on multi-task learning; different vision applications based on the unified framework. Our final aim is to improve the effectiveness and practicability of cross view analysis algorithms with a better solution.

随着社会信息化推进和大数据时代的到来，智能视频分析在社会生产生活中发挥着越发不可替代的作用。跨视角问题是智能视频分析的核心难题，已成为进一步提升算法性能的瓶颈，并限制了智能视频分析方法在多场景海量数据处理中的推广应用。本课题针对跨视角智能视频分析开展系统研究，在归纳现有算法局限性的基础上，拟将面向不同视角数据的视觉任务通过多层次线索关联在同一个学习框架下进行多任务学习，加强不同视角数据的交互和整合，以期充分挖掘多视角的互补信息，提升算法性能。本课题拟在基于深度学习的属性获取和层次化特征构建基础上，分别探讨基于不同线索的多任务学习策略，研究多种线索的多层次融合机制，进而建立基于多任务学习的跨视角智能视频分析系统框架，开展基于该框架的面向多种视觉任务的示范应用。本课题的目标为提高跨视角视频分析的准确性和便捷性，最终提供更为有效、更为便捷的跨视角智能视频分析解决方案。

跨视角问题是智能视频分析的核心难题，制约了算法性能提升的可能性，也限制了智能视频分析在海量数据中的进一步推广应用。在本项目的执行过程中，对基于多任务学习的跨视角智能视频分析方法开展研究，在智能视频分析中具有重要的研究意义与深远的应用前景。本项目以基于深度学习的属性获取和层次化特征构建为突破点，面向不同视角数据的视觉任务通过多层次线索关联在同一个学习框架下进行多任务学习，探讨了基于不同线索的多任务学习策略和融合机制，建立基于多任务学习的跨视角智能视频分析系统框架，在智能视频分析任务中取得了优异的性能。本项目在本领域国际权威学术期刊和会议上发表论文35篇，申请专利2项，搭建平台2套，推广应用技术1项，开展多次国内外合作与交流活动，圆满完成了项目任务书中的各项要求和预期任务指标。