Research On Human Motion Analysis And Key Technologies In Intelligent Video Surveillance

Intelligent video surveillance technologies are the use of automatic video analysis for video surveillance technologies, as an effective means of security are more and more attention. New generation of intelligent video surveillance technologies involved in image processing, image analysis, machine vision, pattern recognition, artificial intelligence and many other fields of study, it is a multi-disciplinary and comprehensive problem, but also a challenging and leading-edge task. Its purpose is to make the video surveillance system intelligent and replace some or all work of people. With the development of performance of computer software and hardware, the intelligent video surveillance systems with a variety of complex background applications continue to emerge so as to stimulate the interesting of the majority of scientists, research institutions and enterprises in the world. Because of its intelligent development, it makes a wide range of applications in all walks of life, its research has a strong theoretical and practical significance. In this paper, for the key issues of intelligent video surveillance, moving target detection, human tracking and behavior recognition were studied.In this paper, the object of study is the human in the scene which is taken with the static single-camera.First, the objects detection methods are studied. Moving objects detection is one of the most important implementation steps and components in intelligent video surveillance, its purpose is to detect and extract objects from video sequences, so that it can identify and track objects.Video images are collected by static camera, at the same time a large amount of noise generated, together with the light change, the branches or water swing and other background interference in video scenes, these noises and interference will have an impact on the detection results, so we must remove noises before detecting objects, this paper first use the Gaussian low-pass filter to smooth the video image, and then use the background subtraction and two consecutive frames subtraction methods to extract moving objects, and compare the advantages and disadvantages of two algorithms. The key of background subtraction method is to set up the background model, whether accuracy of the background model is directly related to the accuracy of the final detection results. According to the statistical properties of images, the Gaussian background model is selected, because the background is gradual change with time, the original background can not accurately detect objects, so using the real-time update strategy to update the background. To enhance the anti-interference and the stability of the background model, not update the former. After setting up background model, with the current frame image subtract background image, and then obtain the objects.Temporal differencing method makes use of the pixel-wise differences between two or three consecutive frames in an image sequence to extract moving objects, because two consecutive frames subtraction has its inherent flaws, when objects move faster, obtained objects will larger than the actual objects, and when objects move slower, objects that will can not be detected, so in this paper, use three consecutive frames subtraction method to detect objects. This method retains wealth information of original images, at the same time also can restrain some noises.The detected object regions by above two methods are expressed with two values, zero and one, then use the morphological filtering method to eliminate the small holes which are in the moving regions, analysis connected domains, and finally obtain the complete object regions. By the comparison of the two methods, background subtraction method can extract a more complete regions of the human body, but its calculation is more complicated, and has a strong interference on the object shadow, in need of special methods to improve the shadow inhibition; Temporal differencing method calculation is simple, but it will appear many holes in the regions, at the same time, the outline of detected objects is often discontinuity, that will impact extraction of features. Two methods have yet to be improving.Second, the tracking methods of human body are studied. The objects tracking is an important part of video surveillance systems. The objects tracking technology is the creation of matching problem based on the relevant features between two consecutive frames images. Objects are described by the status of their own, so the tracking problem is equivalent to solving the status of the object, the process can be realized by estimation theory. The most effective method is the Kalman filter algorithm, but the Kalman filter algorithm is only applicable to linear and Gaussian system, but most of the practical application of systems are non-linear and non-Gaussian, so this paper study two classical algorithms which solve non-linear and non-Gaussian problems, that is extended Kalman filter algorithm and particle filter algorithm, and compare the two methods with mathematical model, for the conditions of changing the linear change intensity, the initial state, the noise covariance as well as the number of particles, the particle filter algorithm is better than extended Kalman filter algorithm.In this paper, a multi-feature amalgamation of particle filter tracking algorithm is advanced for the actual scenes of video surveillance. Because particle filter has the good performance to non-linear systems, here we use this algorithm to track the human body in videos. First of all, set up the system state model, because the autonomy trend of motion objects is generally more obvious, the spread of particle can be a process of random movement, namely, it subject to auto-regressive process (ARP) equation, so this paper uses the second-order ARP model; then required observation model, the model used to measure objects observation vector and the current status of compliance with the level of observations, based on the different applications and information, which can be defined as any reasonable measurement of the match. The most intuitive observations can be gray images, can also be a variety of features that extracted from processed images, such as color, edge, texture, shape and other features. In this paper, the color and edge features of the human body are extracted before set up observation model, the color feature is expressed by color histogram and the edge feature is expressed by edge detection which uses Canny edge detection method, then, calculate the Bhattacharyya distance between the candidate features and reference features, and integrate into the observation model. The tracking process is a prediction and update repeated iteration process through the status model and observation model, thereby, achieving the tracking of the practical application system. Experiments show that the effectiveness of the algorithm, and the particle filter algorithm which integrate a number of features can overcome the missing track and drift phenomenon by using single and be able to adapt to a variety of environments to track human body.Finally, recognition of human behavior is studied. The recognition of human behavior is one of requiring high level attention and challenging research directions in behavior understanding domain, it belongs to advanced treatment. The part of study in this paper apply the template matching method and structure moment of two-dimensional image to identify human behavior, the structure moment has the features of translation, rotation, scale transform invariant, in the method, the density function of moment invariants is the transformation function by the original density function, so that generates a new structure moment. Through the analysis of experimental results can be seen that the method can effectively identify the body behavior, but the method cannot solve the flows of the traditional template matching algorithm, so an effective improvement is necessary for the algorithm.