Several Key Problems Research On Intelligent Monitoring Analysis System

Intelligent video surveillance is one of important research issue across computer visionand several related subjects. Its fundamental problem is how to effectively analyze thesampled data and then draw a conclusion similar to even superior to that by human. So far, theresearches on low-level data processing methods have gained significant progress.Nevertheless, the extraction and knowledge presentation methods of target with high-levelsemantics are yet to be improved. Especially, it needs essential research to effectivelycombine low-level pre-processing and high-level analysis task. Therefore, to study theintelligent video surveillance is of theoretical significance and applied value.The aim of this dissertation is to research the key techniques of abnormal event detectionin intelligent video surveillance. The approaches are addressed to low-level data processing,mid-level data presentation, and high-level semantics analysis.First, this study discusses efficient pre-processing methods including image denoising,video denoising accelaration, and object extraction in video detection.⑴According to thecorrelation among wavelet coefficients across scales, a Bayesian shrinking denoising methodis proposed using zero-like tree structure. It effectively overcomes the disadvantagespresented in conventional methods, such as missing detail, slow speed, and fake edge.⑵Toimprove the sensitivity to parameters of non-local means (NLM) filtering, a NLM usingwavelet energy (NLMW) is put forward with an adaptive filtering parameter. By exploitingthe relationship between wavelet coefficients’ energy and image texture, this method employsregression analysis to fit a parameter selection function using wavelet coefficients’ energy as avariable. An optimal parameter is thus adaptively determined according to texturecharacteristics of image to be denoised. On the basis of NLMW, a novel NLM method usingwavelet energy and segmentation (NLMWS) is discussed, which segments an image intoblocks by the mean shift and assigns a suitable parameter for each block. It improves furtherNLMW’s denoising quality.⑶To satisfy the real-time requirement of video denoising, thisstudy meliorates conventional three-dimensional non-local means method (3D-NLM). Afterthe moving targets are extracted by the moving detection, different update speeds of coefficients are selected for foreground moving targets and background. Through this means,the denosing speed of3D-NLM is accelerated by four times.Second, this study discusses a local feature extraction and representation method by exploiting the topic model and Hough random forest, to improve the robustness of the recognitionsystem to objects with severe occlusion and changing shapes.⑴By introducing the topic model into image representation, this study proposes a latent structure based topic model, calledLatent Patch Model (LPM), according to image’s characteristics. It builds an effective topic layer on images, so as to overcome the shortcoming of visual words in traditional recognition that can only be aimed at a single task, facilitating the implementation of image processing tasks. To verify the effectiveness of the LPM model, we apply it into image denoising and addressa multiple estimation LPM model (MELPM). Experimental results show that MELPM achieves comparable performance with state-of-art excellent image denoising methods. This indicates the LPM model is able to well represent image’s structures.⑵The Hough forest performs well in object recognition, but we have to overcome certain obstacles before we apply it in specific environments. Since the Hough random forest cannot effectively exploit the spatio-temporal correlation and has poor real-time performance, this study presents an optimization methodfor Hough forest. It uses the dual background modeling to extract foreground information which is chosen as the detection area of test image. After carrying out the Hough voting for multiple images, the proposed method fuses a sequence of temporally adjacent probability maps and makes judgments. Experimental results illustrate that our method obtains higher recognitionrate than state-of-the-art methods.Finally, we studied unusual event etection methods on intelligent monitoring system.There are some deficienciesin traditional optical flow abnormal behavior detection method,such as low degree of differentiation,and too sensitive to shooting angle and the distance. Inorder to improving the optical flow method to test robustness,an improved dual backgroundmodeling which includes an adaptive running average background model and HSVbackground model was proposed. The model can reliably extract the motion area. Foregroundwas obtained from video sequences by background subtraction. The motion area was labeledas several regions of interest, and the optical flow features in each labeled region wereobtained using the Lucas-Kanade algorithm. Amplitude based weighted unit energy derivedfrom the optical flow features was defined to measure the anomaly of human activity.Experiments were conducted on various videos indoor and outdoor, and the results werepresented to verify the effectiveness of the proposed scheme. In this paper, our research results will provide certain support for the machine visiontechnology in intelligent monitoring system of effective application.