| With the development of intelligence,as an interactive object,intelligent products are constantly required to actively adapt to the habits of users.However,it is more and more difficult to meet the requirements of natural interaction just through some traditional methods such as simple keyboard,mouse,and touch screen.People need to develop more direct and natural interaction methods.This paper proposes a non-contact interactive method based on Kinect.Taking traffic command actions as research objects,the static 3D motion recognition technology and the dynamic 3D motion recognition technology of human body are studied respectively.Human static three-dimensional motion is to recognize and classify human motion in a single frame of image and generally includes three steps: acquisition of motion data information,motion feature extraction,and motion recognition classification.After collecting action data and constructing action description features,choosing a model fusion method to fuse decision tree and SVM to classify static actions.Dynamic 3D motion can be considered as a series of static 3D motions.The three steps of dynamic three-dimensional motion recognition are the same as static three-dimensional part.After collecting action sequence data and constructing action description features,recognition and classification were completed using an improved BP neural network.The main work of this article is as follows:(1)An interactive method of 3D motion recognition based on Kinect is proposed.This article introduces the working machanics of how Kinect somatosensory device implements interaction in details.Mainly including depth image acquisition technology and skeleton tracking technology.Kinect can directly capture 25 joint points of human bones,which is an important basis for constructing motion description features in motion recognition classification.At the same time,it also reduces the data volume of the collected human motion data from the pixel level to the skeletal joint point level.(2)Research the related algorithms of static 3D motion recognition.Choosing 14 groups of the angle information between human body structure vectors and 6 groups of the ratio of the modulus between key human body structure vectors as motion description features.Then use decision tree and SVM for classification and recognition respectively,and finally choose the linear weighted fusion method and select the appropriate fusion coefficient,so that the average recognition accuracy is improved.(3)Research on related algorithms of dynamic 3D motion recognition.Choosing a series of static 3D motion description features to compose dynamic 3D motion description features.Based on the use of BP neural network for classification,a hierarchical BP neural network of result-driven algorithm is proposed,which mainly used to re-classify some complex and less-distinguishable actions throught another network.By using the improved BP neural network,the average recognition accuracy is improved. |
