Research And Implementation Of Fast Video Retrieval

With the development of internet technology, the image and video technology has been widely used, it has become an important aspect to do fast inquiry, and retrieval. This thesis put an emphasis on the fast retrieval of advertisement logo in the TV videos and the fast retrieval of pedestrian in surveillance video.In terms of the advertisement logo retrieval in the TV videos, this thesis has proposed and realized two advertisement logo retrieval methods based on local feature SIFT(Scale-Invariant Feature Transform)/SURF(Speeded Up Robust Feature).This method includes three steps, in the beginning, extract SIFT/SURF feature of advertisement logo, and collect the video image with even time interval. Secondly, intercept the ROI of video images according to priori knowledge and extract the SIFT/SURF feature. Finally, do the feature matching. As the FLANN(Fast Library for Approximate Nearest Neighbors)is appropriate for the fast-matching of the feature points in the high dimensional space, the matching algorithm based on FLANN is employed in this thesis, what is more, RANSAC(Random Sample Consensus) is also employed to pick up the matching points. This resolution of TV video is 768×460, the result indicates that the accuracy of SURF based method that proposed method in this thesis is 86.3%, this is a little higher than SURF based method. Furthermore, it can basically meets real time need, the retrieval speed based on SURF is twice as fast as the method based on SIFT,In terms of pedestrian retrieval in surveillance video, this thesis has studied and realized a method of pedestrian retrieval based on human’s form and proportion, This method includes two steps: firstly, frame differential method is employed to detect the motion information, threshold of this period is a variant related to the area of outline of people. Secondly, remove the moving substance which is not pedestrian according to the people’s proportion. Finally, detect the pedestrian. Resolution of surveillance video is 768×480, the result indicates that the method proposed in this thesis reaches the accuracy of 85%, the average testing time of per frame is 52 ms, basically, the method can meet real time need.