Human Action Recognition And Camera Motion Estimation In Compressed Video

With the prices of video capture devices reducing, and the gradual prosperity of videomedia, the quantity of video data shows exponential growth. How to make full use of thesedata is one of the most important tasks of video analysis. Video analysis is a hot topic ofcomputer vision research, including content-based video retrieval, object tracking actionrecognition, etc. In this thesis, we focus on two fields of video analysis, human actionrecognition and camera motion parameter estimation. In the field of human action recog-nition, the state of art method pay much attention on pixel domain feature extraction andimprovement of recognition accuracy, however they ignore time overhead during the com-plex process of feature detection and extraction, which cannot satisfy real-time requiredsituation. This thesis directly extracted features from compressed domain, then generatespatio-temporal templates and global histograms to represented actions. Finally, we useboth sparse group lasso and support vector machine methods for action recognition. Experi-ments are conducted on famous action database-Weizmann. The proposed method has greatimprovement in computation speed, although the accuracy is lower than those conductedin pixel domain, it can be widely used in real-time environments. Meanwhile, this thesisalso starts from compressed domain, conducted research on camera motion estimation. Weuse the Maya software to generate video sequences with camera motion information, thenutilize the motion vector stored in compress video to estimate camera motion. We solvethe selection of macro block by support vector machine. At last, we measure the cameramotion parameter estimation accuracy by residual image of motion compensation and themade a comparative experiments with Tan and Weng both on accuracy and time overhead.Experiments show that the proposed method can obtain better results using shorter time.