A Study On Human Action Recognition In Video

Human motion analysis in computer vision involves object detection,tracking and recognition of human activities. Among which, activityrecognition has a wide range of promising applications in securitysurveillance, human machine interaction, entry/exit control, sports and videoannotation, etc. A large amount of work has been done on the active topicover the past decades. However, human activities always have a hugedifference within the class in the real scene, accurate and fast human actionrecognition is still a challenging and very open question.Based on the study of existing technologies and theories, this paperfocuses on innovative methods and theories of human action recognition. Inthis paper, based on the Bag of Words model, we mainly studied the actionfeature representation, feature fusion, action classification model and otherproblems. The main research problems and contribution are as follows:1. Researched and studied action representation and classificationmethods in human action recognition. Specifically, the optical flow feature,contour feature, the spatio-temporal interest points descriptor, bag of wordsmodel and classification model in machine learning are studied carefully. Thepaper also give a review on the action recognition in two aspect: actionrepresentation and classification model.2. We propose a compact representation for human action recognition byemploying human block-based model and local optical flow features. Thecontour descriptor based on block-based model proposed in this paper can beused to represent the movement of human body exactly by dividing silhouette into different blocks. Meanwhile, we propose an accurate and stable opticalflow descriptor for motion information. Finally, our action feature vector isfused by using contour and optical flow descriptor together with motioncontext information. In addition, we adopt the well-known Bag-of-words(BoW) model to obtain the final video level representations.3. Research on the different model in machine learning, a two-levelclassification system approach based on multi-class support vector machineand global temporal information is proposed using histogram intersectionkernel method for action recognition. Our approach is tested on two publicdatasets: Weizmann and KTH. Experimental results show that our approachnearly achieves a100%test accuracy on Weizmann dataset and outperformssome state-of-art techniques on KTH dataset.