Research On Video Topic Intelligence Analysis Based On NDK Detection

Public video is an important intelligence carrier. In the era of big data, video intelligence information has become an important part of public intelligence, thus creating a growing demand for efficient methods for valuable information. In this thesis, online video is treated as physics research unit, and the topic, which is composed of many videos, is treated as intelligence research unit. The goal of this thesis is to obtain the intelligence knowledge of the topic, and provide the intelligence analysts with technical support for quickly grasping the situation.Firstly, architecture for video special intelligence analysis is proposed. Secondly, the related key techniques of Near Duplicate Keyframes(NDK) detection, topic clustering and topic threading are discussed. Finally, the above framework and methods are validated by designing and implementing a OSVAS(Open Source Video Analysis System). The main contributions of the thesis are as follows:1. A frame of video special intelligence analysis is proposed. In this thesis, we divide the research object into four levels: topic, video, shot and frame, in which topic is the unit relating intelligence. The frame, which is based on NDK detection, follows the general video analysis process, drilling down from the bottom to the top. The topic intelligence knowledge allows the analysts to quickly grasp the development trend of events.2. A NDK detection method based on local features is proposed. Generally, the rotation of keyframes is small. Taking advantage of this characteristic, we combine FAST detectors and BRIEF descriptors, both to ensure the detection accuracy, but also improve the efficiency of detection.3. A link-based topic clustering algorithm is proposed. Firstly, we use NDK between videos to calculate visual similarity, and then we get topic similarity considering transitive relation, which effectively strengthens the connection between the videos within the same topic, and improves the topic clustering accuracy.4. A method of topic value model discovery is proposed. The topic value model declares the inner relationship of the videos and the development rule of the topic. It mainly contains three aspects: the evolvement structure, the trend line and the important videos. These kinds of intelligence enable the analysts to grasp the development trend of events, therefore effectively reduce the cognitive pressure of the intelligence analysts.5. The intelligence analysis platform OSVAS is designed and implemented, which validates the effectiveness of the methods, and provides the basis for the application of the research.