Published Paper Details:

Moving object segmentation and classification fromcompressed video plays an important role for intelligent videosurveillance. Compared with H.264/AVC, HEVC introduces ahost of new coding features which can be further exploited formoving object segmentation and classification. In this paper, wepresent a real-time approach to segment and classify movingobject using unique features directly extracted from the HEVCcompressed domain for video surveillance. In the proposedmethod, firstly, motion vector interpolation for intra-codedprediction unit and MV outlier removal are employed forpreprocessing. Secondly, blocks with non-zero motion vectors areclustered into the connected foreground regions using the fourconnectivitycomponent labeling algorithm. Thirdly, object regiontracking based on temporal consistency is applied to the connectedforeground regions to remove the noise regions. The boundary ofmoving object region is further refined by the coding unit size andprediction unit size. Finally, a person-vehicle classification modelusing bag of spatial-temporal HEVC syntax words is trained toclassify the moving objects, either persons or vehicles. Theexperimental results demonstrate that the proposed methodprovides the solid performance and can classify moving personsand vehicles accurately.