融合时空上下文信息的运动目标视觉分析方法研究

Facing the emergence of massive video data, the ability to give the computer automatically analyze the moving object in the scene is one of the core issues and key technologies in an actual application system which taken videos sequences as inputs. Based on the thought that the pixel value of a location in a video frame is strongly related to its historic and neighberhood pixels in the aspects of time and space, this proposal aims to solve the difficulties in a visual analysis system that get rapid and robust object detection and tacking, effective feature extraction and expression, model the object action and represent semantics, etc. Ultilizing the advantages of spatio-temporal context and multi-feature fusion methods, this proposal is to study the visual analysis methods of a moving object in a video sequence by fusing spatio-temporal context. The major research contents include: rapid and robust object detection methods, object tracking methods based on best fitting of an object and spatio-temporal context, object classification methods based on multi-feature fusion and action recognition methods based on 3D convolutional neural networks, etc. These novel methods will be analyzed and verified in the visual analysis system of moving objects. Through the research of the proposal, we wish to strength the theories and algorithms of visual analysis of moving objects and provide theory and technology supports for promoting the development of the corresponding industries such as video surveillance.

面对海量涌现的视频数据，赋予计算机自动对场景中的运动目标进行视觉分析的能力是以视频序列作为输入的实际应用系统中的核心问题和关键技术之一。基于视频帧中某一位置的像素与其历史像素和邻域像素之间存在很强的时间相关性和空间相关性的思想，本项目主要就运动目标视觉分析面临的快速鲁棒的检测与跟踪、特征提取与表达、行为建模与语义表示等问题，综合利用时空上下文信息和多特征融合的优点，研究视频序列中融合时空上下文信息进行运动目标视觉分析的方法，主要研究内容包括：快速鲁棒的目标检测方法，基于目标最佳拟合和时空上下文信息的目标跟踪方法、多特征融合的目标分类方法、基于卷积神经网络的目标行为理解方法等，并将研究成果集成入运动目标视觉分析系统进行分析和验证。通过本项目的研究有望丰富运动目标视觉分析的理论和算法，为促进视频监控等相关产业的发展提供理论和技术支持。

本项目基于视频帧中某一位置的像素与其历史像素和邻域像素之间存在很强的时间相关性和空间相关性的思想，主要就运动目标视觉分析面临的快速鲁棒的检测与跟踪、特征提取与表达、行为建模与语义表示等问题，综合利用时空上下文信息和多特征融合的优点，研究视频序列中融合时空上下文信息进行运动目标视觉分析的方法。本项目主要针对目标检测、目标跟踪、目标分类和目标行为理解等方法展开了研究，并将研究成果集成入运动目标视觉分析系统进行分析和验证。本项目共发表SCI收录论文19篇，EI收录论文7篇，出版专著1部，授权发明专利4项，获得吴文俊人工智能科技进步一等奖1项。本项目的研究丰富了运动目标视觉分析的理论和算法，为促进视频监控等相关产业的发展提供了理论和技术支持。