Event Detection Modeling And Optimization In Intelligent Video Surveillance

This thesis mainly discusses about modeling, application, and optimization of Bayesian Network (BN), also its specified case as Hidden Markov Model (HMM), in Intelligent Video Surveillance System (IVSS). Such discussions include application of presented models in surveillance event detection, modeling and localization of presented models, structure and parameter optimization, and evaluation of detection quality. In recent years, along with further development of and understanding into advanced technologies such as networking, integrated circuit, embedded system, and computer vision (CV), also with urgent desire for reducing cost in surveillance, system efficiency enhancement, and improvement in robustness, IVSS as the next generation video surveillance system is calling up growing research interests both domestically and abroad. CV algorithms based intelligent image processing techniques, as the very core part of IVSS implementation, have been enjoying worldwide research interests. However, although there are a series of CV algorithms that provide satisfactory detection accuracy, high computational complexity becomes a major problem for practical application. For those simpler CV algorithms, many of they fail to offer reliable detection experience. Thus, CV centric IVSS is not ready for practical use, although some of them promise to run in real time with aid of hardware accelerator. As a matter of fact, upon breakthrough research works into stochastic modeling, such as BN and HMM, and upon deeper understanding into cognitive process, researchers become interested in fusion of CV techniques and cognitive process. In spite of the fact that cognitive process lacks strong mathematical support for its effectiveness at the current stage, it is widely believed that, by introducing cognitive process, a system enjoys, but not limited to, the following advantages: 1) Enhancement of robustness; 2) Higher detection accuracy; 3) Reduction in modeling complexity; 4) Lowered data amount for optimization; and 5) Descending computational complexity. This is also proved by various experiment results. In consideration of advantages presented above, we propose to introduce cognitive process into implementation of IVSS, in the hope to build up a system that could provide more accurate, higher speed, and simpler detection experience. Proposed IVSS consists of three main divisions, that is, background segmentation, surveillance event modeling, and model optimization. According to our targets, we have our research arranged into three divisions:First, we investigate background segmentation methods based on CV algorithms. Background segmentation is meant to provide foreground masks and information of Connected Spaces for cognitive process as inputs. In practical IVSS, there exists a great amount of noise, frequent background disturbance, and illumination chances. In order to solve such problems, we propose a novel Adaptive Adjusted Gaussian Mixture Model (AAGMM) as background segmentation method, based on Gaussian Mixture Model (GMM) framework, with introduction of texture feature, intensity feature, and training pixel to enhance robustness. Meanwhile, we adapt an Online Optimization scheme of Expectation Maximization (EM) algorithm to reduce computational complexity of GMM. We manage to achieve reserved detection accuracy while successfully reducing time proportion for CV algorithms.Second, we investigate modeling of surveillance events and corresponding model optimization with two stochastic approaches, BN and HMM. Starting from accuracy requirement, we propose a BN based surveillance event detection (SED) for IVSS modeling, which is practically implemented on Illegal Access (IA) detection to check its practicability and detection accuracy. In the meanwhile, considering that search space for optimizing BN could possibly grow intractably huge, we abandon the traditional way of applying EM algorithm for optimization. Instead, we propose to employ Genetic Algorithm (GA) for BN optimization and investigate practical implementation of GA onto real BN models. In respect to the disadvantage of lacking flexibility of BN, we apply another SED modeling method by proposing a novel Cross Layer HMM (CLHMM) architecture. CLHMM supplements inferences in between neighboring layers and non-neighboring layers based on hierarchical architecture of Layered HMM (LHMM). The purpose of such scheme is to reduce system redundancy, lower complexity of optimization, improve flexibility of modeling, reserve data for optimization with limited training data, and in the meanwhile retain satisfactory detection quality. CLHMM employs Baum-Welch algorithm as optimization approach, based on which we supplement optimization for inferences among different HMMs.At last, in respect to IA detection, we investigate performance of both BN based and CLHMM based models in terms of detection accuracy, false alarm rate, time consumption for detection and optimization, and robustness. Furthermore, we compare the performance by these two methods. Experimental results indicate that both models succeed in achieving detection accuracy as high as CV based schemes, with advantage to be able to run in a real time fashion. From the point of view of system complexity, CLHMM overcomes BN by less expense in learning and higher complexity of modeling. Thus, CLHMM proves stronger practicability over BN in practical IVSS utilizations.