Research On Video Surveillance Data Detection Based On GPU

With the development of technology and the growing of people’s safety awareness,video surveillance system is widely used in people’s lives, leading to the growing of videosurveillance data, and video surveillance data detection attracts more and more attention askey technologies of video analysis. Although there are many detection methods proposedin recent years, these methods cannot meet the requirements of accuracy and real-timeperformance simultaneously. Therefore, it is important to improve the accuracy andreal-time performance of those existing methods, and the main research is focused on thefollowing two aspects in this paper.In aspect of moving object detection, a background subtraction with spatial-temporalGaussian mixture model is proposed based on the deeply research of the existing movingobject detection algorithms. This method improves the traditional Gaussian mixture modelin several ways. It takes into account spatial and temporal dependencies, as well as anadaptive threshold depending on the measure of background dynamics for the foregrounddecision. In addition, this method is parallelized on GPU with multiple optimizationtechniques to achieve real-time performance. One of innovation optimization techniques isthat the local Z-shaped block storage is used to storage the video image data instead of thewarp storage based on the local characteristics of video image.In aspect of keyframe extraction, a keyframe extraction method is proposed based onlocal foreground entropy. The local foreground entropy information is used to measure themotion information in video, and the triangle model of local foreground entropy is built toextract keyframes. In order to achieve real-time performance, a parallel scheme ofkeyframe extraction is given, mainly for the calculation of local foreground entropy. In theprocess of parallel implementation, the mapping granularity of parallel computing isselected reasonably to take full use of GPU computing resources.Experimental results shows that the moving objects detected by the spatial-temporalGaussian mixture model are more accurate, inhibiting the effect of noise, moving shadowsand changes of background. Meanwhile, the keyframes extracted by the proposed methodare better representing the contents of surveillance video. In addition, the parallelization ofmoving object detection and keyframe extraction both gain significant acceleration effect.