Human Action Recognition In Realistic Environment

Now the camera equipment became popular, the speed of internet increased and prices of storage devices kept to fall, which led to the rapid expansion of the video resources. These resulted in lots of important applications:such as video surveillance, video retrieval and human-computer interaction. It is difficult to satisfy the actual requirements of the security monitoring by using manual analysis to video in the field of public safety. So it became very important to use computer to analysis the video, replacing the manually. Because most videos are related with human, recognizing human actions from videos has become an important research topic. Human action recognition can be divided into two steps:motion representation and action classification, in which the extraction the appropriate features to represent the body’s movement is a very difficult problem, especially in the realistic environment. Because of the differences between actions, environment and recording settings variations and rate variations, the representation of the human body movements become more difficult. In recent years, researchers have been concerned with the design of new image representation to characterize human actions, especially the action template which may straightforward integrate with static classifiers such as SVMs or Adaboost. This article will present two new action templates and combine the two action templates to represent human actions.The outline of the paper is listed as following:1. The analysis and expression of body shape in video sequences are researched. Based on the elimination of some background noise, the static edge information of human body was accurately extracted from the video sequences, which was used to construct accumulative edge images, to represent the human actions for action recognition.2. The analysis and expression of human motion in video sequences are researched. Based on the principle of the noise between frames was independent and the action information is related, the dynamic characteristics of human actions was extracted by spatiotemporal filter, which was used to construct accumulative motion images, to represent the human actions for action recognition.3. The dynamic characteristics represent the human motion and the static characteristics represent the shape of human movement, which were combined to represent human action for action recognition. The binding mode was explored in depth.The results on YouTube dataset showed that the performance of accumulative edge images and accumulative motion images exceed motion history images. Moreover, the joint of accumulative edge images and accumulative motion images was better than only one representation, which used in action recognition.