Research On The Method Of Event Database Construction Based On Video Sample Classification

Traditional video retrieval is video-based artificial tagging keyword search, but when the amount of video data is very large, the need for human consumption will rapidly increase. Therefore, this article presents the video incident detection method based on the semantic content of the video event repository established.Based on the main purpose of the semantic content of the video event repository is created for the main event and target semantics of a video identification. Video image consists of a series of frames, and the main body of the target can select several video frames to identify the representative frames, these frames are called representative key frames. SURF descriptor represents a characteristic point of the frame, when changes in the larger picture of the video when the number of adjacent matching feature point between frames will be a sharp decline, this article uses a key one kind SURF descriptor-based matching rate frame extraction algorithm. Video image compared to more continuity in time, but the main goal of the video track data can well reflect the continuity of video over time, thus using the moving object trajectory data extraction usual manner(Gaussian mixture model and Came Shift algorithm combined extract trace data).Track data video key key frame and the main target of no way to direct the video event to identify the key frame can reflect the body of a video moving target object, extract the video’s main target after the underlying feature vector video key frame to make a preliminary classification. The observation that the number and background of key objectives in the video frame complex scenes video key frame that is very different, and therefore adopted based on word frequency vector visual word video scene made a semantic clustering. Track data cannot be directly used to identify the event in the video, so we propose a seven-dimensional vector quantization weights calculated for each section of track data and dimensional vectors. A different number of different video track data is often extracted, and therefore this paper presents a calculation method of similarity between different data tracks the number of video-based. Finally, the video test sample data sets with K-Nearest Neighbor algorithm for video semantic predict when the event is too large library of video, each with a video comparing spending too much time, so using a library then sampling event the video is incomplete, the experimental results showed little difference between the predicted results, but the speed can be greatly improved.