Human Action Recognition Research in the Video Surveillance
|School||Taiyuan University of Technology|
|Keywords||video surveillance motion recognition Camshift algorithm Huinvariant matrix support vector machine|
As the enhancement of people’s awareness of security, video surveillance has been widely used in much more important domain, which brings a lot of convenience for people with the development of video surveillance. Due to the constantly changing of people’s needs, the requirements of the video surveillance are also changed. The intellectualization of video surveillance has become the main theme of the video surveillance development, especially the recognition of video action and the tracking of abnormal motion which now is a focus for researchers. The video is much more sensitive to the variable circumstance, block, distance, light and other factors, which is inevitable during the research. During the recognition and tracking of the video surveillance, these factors severely affected the accuracy of the result. While the development of the video surveillance is to meet people’s needs, and makes those external factors artificially lowered as much as possible to a minimum.This thesis firstly describes the intelligent video surveillance applications which mainly include the-significance of actual application, previous achievements, latest research institutions and existing problems during the research. The purpose of this thesis is to study and identify some actions in the video images, and it can be divided into three parts. The first part is the pre-processing of the video images, which provides detailed information on framing videos, image de-noising and the basis how to select one method of the de-noising. The second part is the detection and tracking of moving target, which analysis the optical flow method, background subtraction and frame difference method as well as their advantages and disadvantages. In view of the requirements of this thesis, the background subtraction method is selected. The part mainly describes the Meanshift algorithm based on tracking and the improved Meanshift algorithm which named Camshift algorithm, at last, one improved Camshift algorithm is proposed, which basically solved the distance and shelter issues. The third section describes the feature extraction and on the characteristics of the training and classification. First, the accumulated motion image from the movement concepts and features to its gray projection. Then, with the introduction of Hu invariant matrix, which Provides detailed information on Hu invariant matrix when the image translation, rotation and scaling invariance principle. Finally, Hu invariant matrix is extracted on that basis and the feature is trained and classified. This thesis provides detailed information on the principles, applications, and characteristics of the support vector machine. In this thesis, we select a widely used support vector machine as classifier.At last, in the Visual Studio2010environment joint OpenCV2.4.3, a software is developed based on C/C++language to test some of the typical video clips, and self-timer fragment. Experiment shows that this software can achieve real-time accurate tracking and identification of some fragments of the video.