The Research On Human Action Recognition Based On The Action Attribute-Classifiers

With the rapid development of multimedia technology in recent several decades, computer vision has a widespread applications in aspects such as military, medical science, manufacturing and so on, which also facilitated related subjects such as pattern recognition, artificial intelligence, machine learning to do further researching. Human action recognition is becoming the research field attracted much attention in computer vision and has a wide range of applications and utility value in military and civilian fields such as intelligent video surveillance, video retrieval, virtual reality, human-computer interaction and so on. Although the researches on human action recognition have got some achievements, but we have no a universal recognition method so far, we are still facing a lot of difficulties and challenges in this field. This paper focuses on the research on single person-action recognition method which is to describe the actions reasonably and classifying them in the video. This paper analyzes the research status and introduces the basic theories. Due to the defect of low-level visual features based traditional action recognition method cannot perform well sometimes, we propose an action attribute based recognition method. The main stages of the proposed algorithm are divided two parts:training action attribute-classifiers and constructing the recognition model. In the stage of training action attributes classifiers, we firstly detect spatio-temporal interest points in the video and extract3D-SIFT descriptors around each interest point. Then we use bag-of-words to describe each video to generate the low-level features samples. Next we label the manually defined attributes for the parts of the low-level features vectors and train SVM attribute classifiers with them. In the stage of building the model, we use SVM attribute-classifiers to predict the attributes for each low-level feature vector without attribute labels and finally the structure and parameters of Bayesian network model is learned via these generated attribute-behavior cases. In the testing phase, we used SVM attribute-classifiers to predict the attributes for the unidentified behavior and then apply the Bayesian network recognition model to classify it. The behavior which has the maximum probability is as the result. The recognition method is validated in Weizmann datasets and KTH datasets in this paper. The experiments results illustrate that our method is more effective than traditional low-level visual features based methods.