Violent Event Detection Algorithm Based On The Spatio-temporal Features Of The Video

With the development of information technology,the popularization of the Internet and the visualization of network data,online videos are increasing sharply.The security of the massive video contents has been widely concerned by the government and the society.Research on the security of video content has an important practical significance.As violent video is a main part of special video,how to effectively represent and accurately detect violent event in the video becomes the main target of this thesis.In this thesis,optical flow algorithm which reflects the spatio-temporal semantics of the video is used to detect the violent event in the shots.There are several innovations in this thesis and the work is organized as follows:1.An algorithm based on motion boundary trajectory of gradient image and extreme learning machine is proposed.Firstly,the gradient method is used to preprocess each frame of the video.Then,the gradient of optical flow field between consecutive frames is calculated.Finally,a dense trajectory algorithm is used to extract spatio-temporal dynamic features.After using Bag of Visual Word to reduce the dimensions of the feature represented in high-dimension,the Extreme Learning Machine is trained and tested by the vectors.The experiment result on the Movies dataset shows that such hybrid model combined with Extreme Learning Machine has a good performance in violent event detection.2.An algorithm based on optical flow and Co-occurrence Histogram of Oriented Gradient is proposed.Firstly,the optical flow field between consecutive frames is calculated.Then,the Co-occurrence Histogram of Oriented Gradient of optical flow is extracted as the spatio-temporal dynamic features in the video.Finally,the features dimensionally reduced by the means of Bag of Visual Word are put into Support Vector Machine for training.It not only performs better on the Movies dataset,but also obtains good result on the more complex Hockey dataset,which represents strong generalization ability.