Human Action Recognition Based On Depth Motion Maps

With the rapid development of computer vision,human action recognition is an important branch.Many universities and companies have begun to focus on it.Because human action recognition has a very important application prospect in the field of public safety,China has invested a lot of money in recent years.As the application scope expanded,how to improve the recognition accuracy of human action and the robustness of the algorithm has gradually become the focuses of the researchers.The emphasis of this paper is to study the characteristics of human action description and human action classification of these two aspects.Because the traditional human action recognition methods is mostly dependent on the ordinary RGB cameras.The recorded human action is subject to the influence of lights,visual angles and other factors.And these can be greatly influence the recognition rate.This is disadvantageous to the commercialization of the human action application.Therefore,this paper attempts to describe the human body by using the depth maps.It can avoid the influences caused by illumination and visual angles.Firstly,we use the Kinect which was developed by Microsoft to collect the human action data.We need the depth maps of human action,and we will use the description advantage of depth maps in depth and texture.Then we use the depth maps to project to three perpendicular of the Descartes plane and converts the depth maps into the desired feature vectors of depth motion maps.Then we use the LBP to extract the features from the depth motion maps again.The formation of the final descriptors of human action is what we required.Secondly,the feature vectors of depth motion maps is reduced the data dimension by using PCA,so as to remove the redundant feature vectors and improve the computational efficiency.Then the L2-regularized collaborative representation is used to represent the human behavior.In the application of collaborative representation,we will use the Tikhonov regularization,the purpose is to adjust the test samples and generated dictionary's weight coefficient.The similar actions will increase the weight coefficient,and the dissimilar actions will reduce the weight coefficient.It is aim to improve the classification effects.Finally,this paper will use Matlab to carry on the simulation experiment,and use the MSRAction3 D database as the standard database of training and testing.In this paper,the MSRAction3 D database will carry on in the condition of Experimental One and Experimental Two.In Experiment One,our algorithm will be compared with other methods,such as Bag of 3D Points,DMM-HOG,HOJ3 D and Space-Time Occupancy Patterns,and the results of the experiment are very good.In Experiment Two,our algorithm will be compared with DMM-HOG,Random Occupancy Patterns and Actionlet Ensemble,and the recognition rate of our algorithm is still better than the other methods.