Research On Human Action Recognition Based On Local Features

Action recognition in video is a hot research topic,which has been widely used in medical treatment,video surveillance,motion analysis and other fields.Although human behavior recognition technology has made some progress at present,the complex and changeable scenes and people's requirements for safety performance bring more challenges to human action recognition technology.How to design a recognition algorithm with both reliability and security has become an urgent problem.In this paper,human action recognition based on different local features is studied to improve the recognition rate of human action and reduce the complexity of the algorithm.The main work and innovation are as follows:1.An algorithm based on multi-scale weighted HOG is proposed to realize effective recognition of human action in videos.In view of the problem that traditional HOG cannot highlight the information of the moving parts in the image,this paper adopts the HOG based on the weight of optical flow intensity according to the characteristic that the larger the motion amplitude is,the greater the intensity of optical flow is.That is to say,the intensity of optical flow between frames is used to weight the gradient amplitude of pixel points,and the histogram of the gradient direction is counted to get the weighted HOG.In order to improve the recognition performance,the weighted operators of different scales are combined in series.Considering that the high-dimensional features cannot be classified directly,this paper adopts the word bag model to construct a visual dictionary and sparse represent the HOG features into a frequency histogram,so as to finally realize the design of multi-scale weighted HOG algorithm.Experimental results show that,compared with HOG algorithm and weighted HOG algorithm,the algorithm proposed in this paper can obtain better recognition effect.The introduction of optical flow weighting and multi-scale fusion mechanism enhances the feature description of the moving parts and enriches the feature information.2.Considering the HOG feature lack of scale invariance and SIFT operator in the scale space have good invariance and stability,in order to improve the performance of the algorithm,this paper further multi-scale weighted HOG and SIFT features fusion,thus making the fusion feature can also contain the target contour information,interest point information and motion information.Aiming at the problems of inconsistent interest points of SIFT and high similarity between classes,this paper adopts K-means clustering algorithm to build visual dictionary,codes the SIFT features into vector histograms of certain dimensions,so as to enhance the differentiability of features,and finally series the acquired SIFT features and coded HOG to get the fusion features.The experimental results on KTH show that the proposed algorithm can achieve a better recognition rate than the algorithm using a single feature,and the two features can play a better complementary role.3.Aiming at the deficiency of human action representation based on two-dimensional pose,an improved two-dimensional human pose feature parameter coding method was proposed to reduce the complexity of posture feature representation.Human action recognition in video involves the extraction of spatio-temporal information of action.In this paper,the spatial information is firstly represented by the Angle features calculated from the angles between adjacent skeletons of the original two-dimensional attitude.The time information is first encoded into the parameter space(?,?)with the Hough Transform,then the pose is mapped to the point in the parameter space,and finally expressed by its motion trajectory.After the representation of spatial and temporal information is obtained,the Bo P model is used to construct the pose codebook,and FV method is used to encode the Angle and trajectory features respectively,so as to obtain the corresponding FV feature vectors.Furthermore,the two features are combined in series to represent the spatial and temporal features of the pose.Compared with the algorithm proposed by Varges,the results show that the classification performance can be improved by encoding the two-dimensional pose into the parameter space,and the fusion of spatial and temporal features can further improve the recognition rate.