Violent Video Detection Based On Spatial-Temporal Features

With the development of Internet videos,people have an easier access to various videos.Among all kinds of videos,there are some violent videos which are harmful to teenagers.It's necessary for us to tag and control this kind of videos to keep them away from them.As for it,manual work are general ways to do this.But the amount of videos are so large that manual work are unable to deal with all the videos.As a result,proposing an accurate automatic violent videos detection algorithm is very meaningful.This work does research on global features and local features and propose three violent video detection methods.First,this work proposes the violent video detection method based on Mean Diverse Density.This work analyze the content character of global features and chooses CSD as the suitable descriptor of global features.This work also analyzes the structure character of global features and finds out part-violence problem which is hard to deal with by traditional supervised learning methods.This work introduces multi-instance learning(MIL)to solve this problem.In consideration of the specialty of violent video detection,we propose a new MIL learning algorithm named Mean Diverse Density(MDD).Experiments show that the accuracy of proposed method is higher than existed MIL algorithms and traditional supervised learning.Second,this work proposes the violence detection method based on dimension reducing Bag of Words.This work analyzes the character of local features and chooses Mo SIFT to extract the video features.We use Bag of words(Bo W)to detect violent videos.To solve the long time consuming of training dictionary,this work proposes length reducing algorithm according to the character of K-means.To increase the accuracy of Bo W,this work proposed width reducing algorithm based on theory of distinguishing dimensions.Experiments show that length reducing increase the proficiency of training the dictionary and width reducing algorithm increase the accuracy.The accuracy of proposed method is higher than the original one.Third,this work proposes the violent video detection method based on spatial-temporal features.This work analyzes the character of global features and local features and proposed the combined detection methods.This method quantizes the results of the first method and the second method and gets a new feature to detect.The experiments show that the accuracy of method based on spatial-temporal is higher than the method based on only one kind of feature and existed methods.