Surveillance Oriented Video Structural Description And Parallel Optimization

With the widespread application of surveillance video systems, thedescription and management of massive surveillance video becomes anurgent problem. Text-based description provides a strong semanticinteroperability, but is weak in processing massive video data;Content-based description eliminates the human assistance in descriptiongeneration, but is weak in semantic interoperability with the existence of“semantic gap”. Video structural description combines the advantages ofboth, provides strong semantic interoperability and automaticitymeanwhile, and becomes the hot spot in the video description field.Based on MPEG-7, this article proposed a general surveillanceoriented video structural description scheme, designed and realized anautomatic structural description algorithm framework, and researched inimproving the interoperability and timeliness of the framework.Firstly, to reduce the redundancy and vagueness of MPEG-7description, a profiled surveillance oriented description scheme wasproposed, after analyzing the characteristics of surveillance videos. Theproposed scheme is compatible of MPEG-7, semantic and object-oriented,and solves the commonality problem of traditional description schemes.Secondly, based on the proposed scheme, an automatic surveillancestructural description algorithm is designed and realized, and reduces thehuman interactions in traditional description systems.Then, this article researched the implementation of the semantic feature extraction module of the algorithm framework, and solves theinteroperability problem. In extraction of color semantics, color mappingcharts are built to extract the semantic color names of objects, which isthe basis of object retrieval; in extraction of texture semantics, this articleresearched the extraction of TBD and object texture types, which solvethe texture semantic labeling problem; in extraction of shape, an SVMclassification method was proposed to extract the shape types, andimproves the interoperability of the low-level shape features.Lastly, a multi-layer parallel optimization algorithm was proposed tospeed up the serial structural description algorithm. In the top-layer, apipeline algorithm was proposed; in the mid-layer, a functionality dividealgorithm was proposed; in the bottom-layer, a data divide algorithm wasproposed. The proposed parallel algorithm utilized the CPU resource ofthe multi-core platform, got an average speed-up of3.5, and guaranteedthe timeliness of the structural description algorithm (30fps).