Research On Human Action Recognition Based On Skeleton Information In RGBD Video

Human action recognition is an important research hotspot in the field of computer vision.It is widely used in intelligent video surveillance,motion behavior analysis,human-computer interaction,virtual reality and so on.In recent years,with the appearance of Kinect and other somatosensory cameras,the focus of action recognition gradually shifted to RGBD video.The RGBD video can provide the information of depth and skeleton which makes the problems of the video sequence based on video sequecnces,such as complex background,light effects,etc.,can be resolved.Based on the review of human action recognition in RGBD video,this paper analyzes the defects of human action recognition in RGBD video,and makes the following two parts according to the current shortcomings.(1)Adaptive feature selection method based on joint feature entropy.This method selects two characteristics,which are HON4D and the relative distance characteristics of joint.According to the random forest error rate of HON4D,the HON4D characteristic of the joint with high discriminant is selected.The relative distance feature of the joint points calculates the relative distance according to the other reference points of the joint.The most compact feature information is retained on the feature,and the time precision is extracted using the Fourier Temporal Pyramid.In the feature entropy calculation,the average information entropy on each joint is calculated using the voting result in the HON4D training model as the adaptive entropy.According to the number of entropy in the test process,whether the HON4D feature of the sample is representative,otherwise the relative distance feature of the joint is used.This method not only solves the increase of the feature dimension caused by the simple splicing,but also solves the problem that the two features are linearly indistinguishable and the result is only affected by the other characteristic.The experimental results verify the effectiveness of the method.(2)High discriminant human joint selection and feature representation.The method uses RGB video as the main feature.The HOG3D characteristic of the joint motion trajectory,which is the HOG3T characteristic.This feature can better extracts the relationship between time and behavior,reduces the complexity of feature extraction and the redundancy after feature extraction.Compared with the original characteristics of HOG3D,the classification accuracy is obviously improved.In order to further reduce the feature dimension,and to understand the realization of behavior in semantics,this paper argues that the main behaviors of action are concentrated in large changes in the movement of the joints.Therefore,this paper puts forward the accumulation of joint movement as the criterion of the joint discrimination.According to the sort of the joint point discrimination force,human action recognition can be classified.The experimental results show that the recognition accuracy of HOG3T based on high discriminant force is improved compared with the original HOG3D-based recognition method.