Action Detection Based On Deep Learning

In the field of video analysis in computer vision,the action detection has two main tasks:temporal action detection and anomaly action detection.Temporal action detection is to detect segments that contain action from an untrimmed video,and gives the timeline location of the action segment,in other words,gives the start and end moments of the action.Abnormal action detection accurately finds those sparse abnormal actions in videos that are almost entirely normal action.These two tasks are the crucial research topics in the field of video analysis.It is of great significance to achieve fast temporal action detection and effective abnormal action detection for the purpose of video content automatic analysis and video understanding.Therefore,this paper proposes a fast temporal action detection method based on action subject detection and an abnormal action detection method based on depth multiple instance learning ranking model by using deep learning technique.The main works of the paper are as follows:(1)As for the action detection methods based on deep learning,the feature extraction parts are mainly researched on.Their advantages and disadvantages are analyzed for subsequent selection and improvement.In addition,three typical action detection methods are implemented,and small improvements are made.The advantages and disadvantages of them are quantitatively analyzed,and the improvement direction of the action detection algorithm is found.(2)A new fast temporal action detection method based on action subject detection is proposed.Aiming at the detection speed problem of existing temporal action detection method,this paper analyzes the propose candidate area module of the detection model and the surveillance video characteristics.Firstly,a pre-action subject detection module filters out non-action segments to quickly generate proposal areas.Secondly,an efficient action/background classification network is adopted to determine whether a proposal area is an action.Finally,the timeline location information is the output.Experiments which verify that this method can achieve fast temporal action detection.(3)A new anomaly action detection method based on depth multiple instance learning ranking model is proposed.Aiming at the problem that current video abnormal action detection methods are not effective,this paper first analyses the characteristics of abnormal action in surveillance videos and the weakly-supervised video annotations.By improving the multiple instance learning method and the two-stream network model for video analysis,the accuracy of anomaly action detection is effectively improved and the false alarm rate is reduced.