Research On ROI Extraction And Target Detection Technology For Surveillance Image

At present, the monitoring models based on static single-frame image are mainly used in the fields which have wide distributions, less real-time demands and do not need continuous monitoring. And, the ROI extraction and goal detection in the monitor image is the key to realize the intelligent surveillance. Compared with the monitoring models based on continuous video sequence analysis, because the static single-frame image does not include the motion information of target and the monitoring scenarios are different, the traditional target extraction and detection methods based on motion detection are difficult to apply, thus the hardness of the ROI extraction and detection in monitoring scene is increased. Therefore, research on ROI extraction and target detection technology suitable for the changes of scenes has important practical significance and theoretical value.Aimed to the monitor mode based on static single-frame image, centered on the segmentation of key monitoring areas in monitoring images and the ROI extraction and target distinguish, this paper has researched a lot from these three aspects to obtain more reliable and practical approach for intelligent monitoring.In the aspect of key monitoring area segmenting, with the characteristics of key monitoring area, a segmenting approach based on improved edge detection is proposed. Firstly, remove the effects of illumination, stage lighting etc. on the key monitoring areas through leading into rapid wavelet transform. On this basement, an effective sample edge pixels in the key monitoring areas is proposed and evaluation methods of sample edge pixels are created according to the conception of monitoring centre area, and therefore those forged sample edge pixels are removed. Finally, the edge of the key monitoring areas is obtained using least square method fitting the edge pixels in those areas. In these procedures, computations are less and exactness is raised, so the key monitoring areas of images can be more exactly extracted.In the aspect of ROI extracting, aiming at the problem that the ROI can not be correctly extracted base on the color diversity method when the differences between the noise and the background are more than those between the prospected goals and background, a ROI extraction method is proposed based on the background probability statistic model to separate the prospects and backgrounds of the key monitoring areas. On this base, the accuracy of ROI extraction is increased through controlling background noise with block statistic techniques.In the aspect of target detecting, with the feature of extracted ROI, target feature-based detecting algorithm is proposed. By selecting different contour features of targets, the contour feature detector based on fuzzy C means clustering algorithm is established to distinguish human bodies from objects. Furthermore, according to the color differences of ROI, the color feature detector is presented to distinguish correctly the human bodies from objects.At last, combined with the key surveillance area segmentation, ROI extraction and multi-object detection algorithm, the paper has carried out the experiment using the typical indoor and outdoor surveillance scene images as experimental data. The experimental results have certified that the algorithms can effectively segment the key surveillance area, extract ROI and correctly make a distinction between the human body and objects, thereby improve the intelligent monitoring of static images.