项目名称: 动态场景下视觉事件建模与识别方法研究
项目编号: No.61272251
项目类型: 面上项目
立项/批准年度: 2013
项目学科: 自动化技术、计算机技术
项目作者: 张丽清
作者单位: 上海交通大学
项目金额: 80万元
中文摘要: 视觉事件识别指的是从在自然视觉场景获取的视频中对发生的事件进行识别和描述。该问题不仅涉及到计算机视觉特征分析和模式识别问题,而且涉及到高层视觉认知表征问题,具有重要理论意义,同时具有广泛的社会需求和应用价值,在远程视频监控、智能交通、视频检索和新一代人机交互中有着广泛的应用前景。本项目研究动态场景下视觉事件的建模与描述方法,利用主动视觉机理选择视觉场景与事件主体相关特征,在运动描述空间上刻画视觉事件。进一步在视觉运动特征空间上建立事件的动态贝叶斯描述模型和模式识别方法。本项目研究主要内容主要围绕三个层次上的问题进行研究,包括局部特征描述,事件运动模式描述以及动态模式聚类与识别。在理论上分析识别算法的计算复杂度和计算效率,通过视觉事件建模的典型应用验证提出方法的性能,展示提出事件建模方法的优越性。研发视觉事件建模理论的一个典型应用,为视觉事件识别提供技术原型和相关的实验测试数据。
中文关键词: 视觉事件检测;张量稀疏分解;层次化概率模型;异常事件检测;深度神经网络
英文摘要: Visual event recognition is to recognize and describe the event from video taken from natural visual scenes. The problem covers from low level feature analysis to high level visual cognition, thus is of great theoretical significance, practical value and social demands. It has broad application perspectives in remote video surveillance, intelligent transportation, video retrieval and human computer interface. The project investigates the modeling and description of visual events in dynamical scenes. We employ the visual selective attention mechanism to find event related regions and characterize visual events by using motion futures. Then we formulate the dynamical event models in the motion feature space. The technical approach is implemented into three stages: local feature extraction, event description models and dynamical event clustering and recognition. Theoretically, we are to analyze the computing complexity and recognition performance of the proposed methods. Typical prototypes and computer simulations of visual event recognition will be provided to show feasibility and performance of the proposed methods.
英文关键词: Visual Event Detection;Tensor Sparse Decomposition;Hierarchical Probabilistic Model;Abnormal Event Detection;Deep Neural Networks