A Parallel Metadata Extraction System For Moving Objects In Surveillance Video

With the promotion of projects like Peace City, Safe Campus, etc., the number ofsurveillance videos increase dramatically. Using human resource to handle the video cannot reach the requirement. In order to better analyzing the surveillance videos by usingcomputer, metadata of the moving objects, including size, location, color and some otherinformation, should be extracted and stored. Therefore, the extraction of metadata formoving objects has great importance. Currently, metadata extraction of moving objectsin surveillance videos relies on the traditional algorithm o f moving object tracking, andthe research mainly focus on the issue of how to improve the precision of metadataextraction. However, as surveillance video is characterized with long surveillance timeand large data size, it will result in low efficiency of metadata extraction for insisting onthe traditional serial iterative algorithm, which will influence the subsequent processing.A parallel metadata extraction system for moving objects is proposed to improve thespeed of metadata extraction. Firstly, split the video to make every video segmentrequire the same time to extract metadata. However, there exists the problem of lattermetadata extraction depends on previous metadata. To solve this problem, metadata isintegrated on basis of the moving objects appearing mostly in two segments. Differentobject keywords are set based on the relationship among video segments. Then extractfeature to establish the distance matrix. At the same time, calculate overlap area to solvethe occlusion problem specific to surveillance video. Using the above method, parallelmetadata extraction system can be realized to accelerate the metadata extraction.Based on Linux and Hadoop, parallel metadata extraction system for moving objectsin surveillance videos was completed through C++and openCV library. Theexperimental result shows that the precision of parallel metadata extraction system ishigher than81%when the precision of serial metadata extraction reach90%. The systemgreatly improves metadata extraction speed at the same time, and the speed-up ratio isrelevant to both the number of video segment and that of computational node.