Research Of Human Action Recognition Based On Composite Spatial And Temporal Feature

Human activity recognition has become one of the most active research topics in theartificial intelligence and pattern recognition field recently, due to its wide applications. Thekey technologies for human activity recognition, including the spatial-temporal pointdetecting, activity representation, as well as the activity classification technology are allthoroughly studied in this paper，this paper proposes the speed method of spatial-temporalpoint detecting, constructs the activity recognition model based on bag of word model andtopic model. The main contribution of this thesis can be concluded as follows.(1) Spatial and temporal feature point detection is the basis of human action recognition,spatial and temporal feature point detection efficiency and the number of detection willdirectly affect the efficiency and accuracy of behavior recognition. In this paper, thespace-time feature point detection method is established in the two-dimensional fast featurepoint detector on the basis of the local neighborhood of pixels segmentation threshold andthen divided two parts with the threshold value of the local neighborhood. First, according tothe principle of the binomial probability distribution, the ratio of the number of pixels in thetwo parts can be calculated; Finally, three-dimensional space non-maxima suppressioneliminates redundant point to find the exact location of the point. The experimental resultsshow that this article spatial temporal feature point detection algorithm has a higher detectionrate, and can be stably extract a sufficient number of feature points, has a lower degree ofredundancy.(2) Space-time feature extraction methods affect the accuracy of the behavior described,behavioral feature extraction method limited to a two-dimensional,feature in the spatialdomain and optical flow feature in time domain can be fused as a composite feature todescribe the different types of behavior. Its spatial distribution not only describes the featureof the local behavior, also consider the behavior of the motion information. Experimentalresults show that spatial and temporal feature has a high behavior recognition accuracy.(3) Recognition of human action model is built on the bag of words model with potentialdrichlet distribution model, bag of words model is built by clustering algorithm based on the extracted behavioral feature, and then feature vector for each behavior description can becomputed with behavioral feature dictionary, and finally the feature can be inputted potentialdirichlet distribution model to train behavior recognition model. Experimental results showthat the behavior recognition model has a higher recognition accuracy and higher efficiency ofbehavior recognition compared with the SVM behavior recognition model.