| As a valid clue to visual understanding, contextual information is significantly applied in intelligent information processing and provides an efficient point of view to improve performance of pattern recognition task, therefore context modeling especially semantic context, spatial context, scale context and temporal context of video are widely researched.In this dissertation, we firstly introduce sorts and applications of contextual information. Aiming to better utilize context for object recognition, we associate relative relations of pairs of categories to capture more context information to alleviate disturbance of semantic shift and bias of spatial context. Also, we model temporal context via cross correlation coefficient of foreground state in inside and outside window to extract the key frames of surveillance video. The main work is as follows:(1) We summarize the research status of context and discuss methods of multi-context interaction. Also, we introduce modeling methods among different contexts. Specially, we analyze spatial constraint in static images, temporal-spatial constraints in dynamic video, and high-level semantic for video analysis.(2) We propose a method for object recognition of multiple categories via semantic and spatial context. Concretely, we establish semantic relations among different categories via Latent Semantic Analysis, and model position and scale context of objects pairs. Experimental performances show the effectiveness of the relative context.(3) We extract key frames of surveillance video via temporal context of main objects. According to cross-correlation coefficient of foreground pixels state in inside and outside region, we obtain the time when objects enter or leave the scene. By evaluating relative velocity and direction of object entering/leaving the scene, we get key frames of video. By means of detection and results provided by background subtraction, we obtain object proposals and understand content of video. |
PDF Full Text Request