Research On Human Behavior Recognition Based On Skeleton Information

People deal with things and express emotions by behaviors,and human behavior recognition has a very wide range of application areas that have not been fully solved.Therefore,in recent years,the problem of human behavior recognition has gradually developed into the field of computer vision.On the one hand,many previous studies are based on RGB video or image information.The disadvantage is that the datum of image is insufficiently robust against complex backgrounds and changes in human scale,viewing angle,and movement speed,which makes the recognition results less accurate or the chemical ability is not strong.On the other hand,the development of Kinect cameras has also promoted the expansion of behavior recognition methods to a certain extent,and people began to pay attention to the advantages of skeleton information in characterizing motion.However,the problem is that such methods rely more on the extraction of bone nodes by the device.In summary,in order to avoid such problems effectively,this thesis uses the current popular bone key point detection method based on video to extract features from bone information for human behavior recognition.Main tasks as follows:Firstly,based on the research in this thesis,behavioral data sets are collected from the collection.The collection method takes into account the complex background,lighting background and multi-viewing factors.It contains nine types of daily human behaviors,and has high research value.The thesis evaluates OpenPose and AlphaPose,which are the most popular bone key point detection methods,and compares their advantages and disadvantages,and combines with research needs,selects OpenPose as the bone key point detection method in this thesis.The skeleton model of the target in the data set is built by OpenPose and the category label is added for further research.At the same time,in order to reduce the influence of irrelevant factors on the recognition accuracy,a four-step preprocessing operation is performed on the skeleton information,including scale normalization,deleting some nodes,deleting invalid frames,and adding missing joints.The above work makes it possible to use the human skeleton for behavior recognition without resorting to Kinect camera equipment.Secondly,this thesis proposes a multi-dimensional skeleton feature that combinesstatic and dynamic features.Considering the relative stability of the local position of specific actions firstly,spatial position features and joint angle features are proposed,which can reduce the impact of scale changes and direction changes.On this basis,pay attention to the periodicity of the movement,and extract features for the entire movement process,including joint speed features,joint angular acceleration features and body speed features.In order to reduce the computational complexity,the principal component analysis method PCA is used to reduce the dimensionality of the features.Set up a control experiment,use k-nearest neighbor,support vector machine and deep neural network for model training and testing.The experimental results show that the fusion skeleton feature based on the 2D shutdown position proposed in this thesis can achieve a recognition rate of 91.7% in the homemade dataset,compared with the recognition rate for a single feature has improved to a certain extent.At the same time,it was verified in the public data sets Weizmann and KTH,and the recognition rates reached 87.6% and 93.3%,respectively.It was verified that the fusion skeleton features can better recognize human behavior and have general applicability in big database.