Research Of Human Action Recognition Based On Video

Human action recognition based on video is an attractive research topic in computervision field, which has made great progress in recent years. And a growing number ofachievements have been applied in the real life, such as intelligent monitoring,human-computer interaction, virtual reality and so on.Most of traditional algorithms to represent human action information are based onthe background cut-off and target tracking. These methods are susceptible to beinfluenced by complex background noise, lighting changes, shadows and occlusion. Tosolve this problem, a method of combining3-D Harris corner points and3-D HOG localspatio-temporal descriptors is proposed to represent human motion features in thedissertation.3-D Harris corner point can provide a compact representation of video dataas well as exact location for spatio-temporal events, while3-D HOG descriptor canreflect the motion information in the whole range. Finally, to represent human action,the word frequency histogram is created for each video sequence through bag-of-wordsmodel.K-means clustering algorithm is used to build a visual vocabulary in the dissertation.To solve the problem that clustering result varies with different initial cluster centers inclassic K-means clustering algorithm, a method to improve the selection of initialcluster centers is proposed. For a specific sample collection, set a suitable threshold Twith the principle to maximize the distances between initial cluster centers. Withmeeting the condition that the distance between any two initial cluster centers is greaterthan T, iterative process is conducted. Experiment shows that the clustering results ofimproved algorithm can reflect true difference among clusters and are more stable.The k nearest neighbor classifier and support vector machine are applied to classifythe extracted human action features. And the performance of two classifiers isevaluated.Experiments of human action recognition are carried out on two video libraries ofWeizmann and KTH. Leave-one-out cross-validation is performed for Weizmann, whileK cross-validation for KTH since KTH video library has a large number of actionvideos. Experiment results show that the proposed method can obtain a betterrecognition rate.