A Violent Video Detection Algorithm Combining YOLO And ConvLSTM

In recent years,due to the trend of Internet video traffic,Internet video content presents a hybrid development,and all kinds of violent and illegal videos are full of it,which has a serious impact on social public security.Nowadays,the number of multimedia videos is increasing with each passing day,and the semantic types involved in violent videos are various,which bring great challenges to the detection of violent videos.How to detect violent video quickly and accurately has become a research hotspot in the field of computer vision.At present,deep learning has achieved great success in speech recognition and image processing,showing a wide range of application scenarios,so it is of great research value to apply deep learning to violence video detection.The main research work of this paper is as follows:(1)The feature extraction of violent video detection method based on machine learning is time-consuming and the video time and space information are less used.In this paper,a violent video detection algorithm combining YOLO and ConvLSTM is proposed,which can better solve the problems of time-consuming feature extraction and less use of temporal and spatial information of video,and improve the detection performance of violent video.(2)Aiming at the phenomenon that darknet-53 network in YOLO algorithm destroys image spatial structure and image input size is fixed,this paper proposes an improved darknet-53 network,which can retain image spatial structure well,solve the problem of fixed image input size,and further improve the feature fusion ability.(3)Aiming at the problem that the traditional ConvLSTM algorithm has many tuning parameters and is easy to overfit during training,this paper proposes a ConvLSTM algorithm based on global average pooling layer.This algorithm reduces the spatial parameters to make the model more robust and has better anti-fitting effect.(4)The thesis experiments were trained and tested on the common dataset of Hockey,Violent flow,Movies,and the latest rwf-2000 video set.Compared with several advanced violent video detection algorithms,the experimental results show that the violent video detection algorithm proposed in this paper,which integrates YOLO and ConvLSTM,has high detection accuracy and better generalization ability.