本项目拟针对视频中特定行为的检测与识别问题，研究基于人体和运动注意的弱监督方法。在弱监督条件下，训练样本中只给出视频级别的类别标记而缺乏检测框、行为的起始和截至时间等标注信息。首先，我们提出使用时空反卷积网络来提取行为的时序信息，为行为的密预测奠定基础；其次在网络训练中，提出基于人体和运动注意的方法来指导网络的学习，使得网络能更加快速准确地将注意力集中在目标区域；然后使用基于梯度的类别激活图方法计算出行为在空间上的概率分布图，从而得到行为的空间位置；最后，针对行为的时序检测问题，提出增加稀疏约束的方式来有效选择具有目标行为的片段。项目将为安防监控视频、直播平台与网络平台视频中暴力、色情等特定行为的检测识别提供技术支撑，促进社会的安全和谐和网络环境的净化。

In this project, we explore the actor and motion based attention for action localization in untrimmed videos, in a weakly-supervised manner. Under the condition of weakly-supervision, training examples are just provided with video-level label but without bounding boxes, start and end frame. To achieve this goal, we propose to apply spatiotemporal deconvolution to extract temporal information, which can be prepared for action dense prediction. Then we propose to apply actor and motion supervised attention to guide the model learning. By this means, it can help the network can focus on the target area more quickly and accurately. The we apply grad based class activation map (Grad-CAM) to calculate action localization map, from which we can localize the action in spatial domain. Then we design a sparse loss function to selection the segments for action temporal detection. By this means, we can apply some public datasets which are just given action labels to detect some specific actions, such as violence, aggravate damage. We can also detect some bad behavior such as pornography in live web and videos, and further provide health security in network environments.