Research Of Human Action Recognition Based On Video Surveillance

Since the end of last century, due to the development of computer vision and the more attention of public security, intelligent video surveillance technology has been widely concerned and researched, and becomes one of the hottest research points. Computer vision research is the use of computers to replace huma n analysis and processing of video images or multidimensional data, including pictures and video sequences, etc., and extract useful information.Human action recognition is a hot research topic but difficulty in the field of the computer vision. The key task is extracting related information from the video sequence, on the basis, the body and the target region are detected. By exploiting a variety of video processing techniques to extract feature the computer can describe all kinds of human action and carry on the analysis, which finally can be used in action recognition. So there is important theoretical and practical value to carry out research on human action recognition, especially when it has been widely applied in intelligent video surveillance.The approach builds on recent work on action recognition are numerous, but the complexity of human body structure and the difference between individuals lead to many difficulties of action recognition. There are lots of recognition models that have been proposed, which usually suffer from less biological plausibility. In order to improve the accuracy, acceleration and velocity, we propose a method for human action recognition by modeling the visual perception system.Visual information in the brain travels along a certain pathway, which passes through the Lateral Geniculate Nucleus(LGN) and finally to the Primary Visual Cortex. In this paper, we simulate visual perception system, mainly done the follow works:Action recognition system consists of three major steps: video surveillance pre-processing, then feature extraction, and finally classification. This article explores the possibility of using gray scale video image as input, and doing simple morphological processing.This thesis simulates the feature detection method of primary visual cortex, as well as the classical receptive field(CRF) of simple cell based on hierarchical feed forward architectures, by alternating max pooling and template matching obtained increasing complexity and scale and position invariant features. At the first layer, Gabor filters to process the video sequence at all position and scale of multiple orientation. At the second stage, down-sampling performed by computing a local max. Template matching at the third stage, and finally the features is obtained by computing a global max. In order to approach a better effect, we proposed a support vector machine(SVM) and hidden markov model(HMM) hybrid classifier model.We test the approach on publicly available dataset(Weizmann), and the results show that our algorithm is less amount of calculation, higher speed, while good recognition effect.