Recognizing Scene Categories In Video Analysis

Video analysis is an important branch in Computer Vision research,extracting semantic information from unstructured video data is alsoa breakthrough in artificial intelligence area. Scene and event arekey components of semantic information in videos, the former onereflects “background” which is comparatively solid, while the latterone describes human behavior. In some military use like Unmanned AerialVehicle （UAV） Investigation and Aviation Photography Analysis，automatic video analysis, especially real-time scene detection holdsa great significance. In some other big data related area, like RoboticNavigation, Satellite Image Processing， scene detection is a keypreprocessing step to further analysis in deeper semantic information.In some Top-Down procedure, such as Special Purpose Visual System,scene detection plays as a semantic constraint in further informationdigging.Through analysis of key frames, especially based on our proposedperceptual manifold modeling method, we build a classifying systemaccording to intrinsic visual principle. This paper gives a detailedillustration of theoretical procedure. Experimental results reveal thetopographic properties of this method. Perceptual Manifold Modelachieved a high precision and advantage in real-time property.Moreover, this paper realized an Aviation Photography Analysis System,involve complete procedures of video processing: decode and transcode, video shot capture, key frame extraction, scene classification. Realaviation videos are applied to testify the feasibility and validityof every block of the system. More experiment based different kind ofdata are carried out to discuss the performance in main circumstances.