The Research On Human Action Recognition Based On The Fusion Of Slow Features

As the technology of computing and networking is being developed rapidly, multimedia inform ation has reached aspects of daily life. Frequent interaction, massive data volume and migration to cl oud service are the three features of the advance of multimedia technology. As the media that carries text, picture and audio, video plays an important role in massive data volume. Therefore, to meet the demand of intelligent analysis of large sized video, research and application in involving field becom e essential. Human action recognition, a branch of computer vision technology, is a promising appro ach to video surveillance, video analysis and human-computer interaction. The central problem of hu man activity recognition is the extraction of features, which in turn is one of major factors that shape the results. The 30 years’efforts to the subject of feature extraction made by the collages and institu tions home and abroad,have achieved fruitful accomplishment, and pushed the study to further level.This paper studies a human action recognition method based on fusion of slow features, whose aim is to recognize individual and multiperson activity through extracting slow features and using feature fusion methods including low-level feature combination and high-level feature fusion. This article describes the basic theories and common methods of fusion of slow features, including slow feature analysis (SFA) learning strategies and feature fusion methods, and proposes an individual and multiperson action recognition approach based on fusion of slow features, the main stages of the proposed approach are divided two parts which are slow features analysis and slow features fusion.In the stage of slow features analysis, firstly, histogram of oriented gradient (HoG), histogram of flow (HoF) and histogram of scale invariant feature transform (HoSIFT) are extracted to capture local information. Then slow feature functions of each kind of feature are learned by supervised SFA learning strategy. According to the learning functions, slow features of HoG, HoF and HoSIFT are computed by method of accumulated squared derivative (ASD) to express temporal and spatial information which changes slowly in the action video. Eventually, the slow features have been formed.In the stage of slow features fusion, low-level feature combination and high-level feature fusion are utilized to fuse the above-mentioned slow features, respectively. For low-level feature combination, a new feature vector is formed by concatenating the slow feature vectors of HoG, HoF and HoSIFT, and then the feature vector is send into support vector machine (SVM) for classification. For high-level feature fusion, we split the slow feature vectors into a learning set and a test set. Cross validation accuracy parameters, which are learned by training the learning set, are weights of high-level feature fusion. Finally, the high-level feature fusion classification result is computed by the weights and classification results of each kind of slow feature using a multiclass SVM.In experiment part, the method based on slow features fusion is validated in Weizmann data set, KTH data set, UT-Interaction data set and CASIA data set, respectively. The experiments results illustrate that the proposed method boosts the performance of individual and multiperson action recognition, and it has potential application value in video analysis and safety early-warning.