Motion Feature Based Video Retrieval

With the development of computer technology, plenty of multimedia information appears on internet. Based on much multimedia data, people propose the content based visual information retrieval technique. Content based video retrieval (CBVR) is a new means to query similar video data. Video data have a unique feature that other multimedia don't have. The unique feature is the motion feature. So it's significant to improve the video retrieval system's performance by using motion feature to retrieve.In this thesis, we research on the motion feature based video retrieval techniques. According to the characteristic of digital video, the motion in videos can be separated to 2 kinds, global motion and local motion. Global motion is caused by camera motion. It's the whole movement of all pixels in scenes. Local motion is the movement of the object in the scene. We study two algorithms in this thesis, which are global motion feature based video retrieval and local motion feature based video retrieval. In the end, we propose a new algorithm which is the vector quantization (VQ) and global motion feature based video retrieval. The contents in this thesis are as follows:Firstly, this thesis systematically summarizes the research status of content-based video retrieval. The motion feature based video retrieval techniques are mainly introduced.We study an algorithm of the global motion feature based video retrieval, and conduct an experiment to this algorithm. First video sequences have undergone a short-term global motion analysis, which will accurately extract the bilinear motion model of the global motion. The eight extracted model parameters are used as motion features for retrieval, and the retrieval scheme is based on the feature point sequential match technique and the calculation of square sum of global motion vector distance. Experimental results show that in certain application domain, such as sports videos, the retrieval algorithm could achieve a near-semantic content retrieval.We study an algorithm of the local motion feature based video retrieval, and conduct an experiment to this algorithm. Based on parameters estimated from global motion model, local object motion information are extracted from general video sequences. The extracted local motion vector field is then used to produce a directional histogram, which is taken as the feature of the video retrieval system. Experimental results show that this local motion information retrieval system could effectively find clips that better match people's understanding of the video's motion content.Vector quantization (VQ) is an effective lossy coding technique. The merit of VQ is the high performance of the code rate-distortion. With the perfect development of VQ technique, more and more scholars are interested in the VQ based multimedia information retrieval algorithms. As VQ technique could map plenty of vectors on limited code vectors in the way of the least distortion, thus vectors quantized could denote the features of plenty of vectors. This assures the effect of video retrieval. The histogram of code vector indices could be used to retrieve motion vector fast and assure a real-time video retrieval.A VQ and global motion feature based video retrieval algorithm is proposed in this thesis. We innovate upon a VQ based algorithm to eliminate the singular points of the motion vectors in the frames. This algorithm utilizes the motion vector index histogram of all frames in the video sequences to match the video sequences. The experimental results show that this algorithm could effectively extract the statistical characteristics of the global motion feature. The calculation complexity is low, and the retrieval efficiency is high.