Human Action Recognition Methods Research Based On Depth Information

Human action recognition is an important research topic in computer vision. It has been applied in security surveillance, human-computer interaction, etc. With the popularity of depth sensor, depth information is gradually applied to human action recognition. In this thesis, global representation and local representation methods of human actions using depth information were studied from two aspects. Then actions recognition was realized using multi-layered motion history images and spatio-temporal interest points based on depth information, respectively.The main work of this thesis is as follows:(1) The key technologies and related work of human action recognition were stated. On this basis, the methods of human action recognition using depth information were analyzed deeply and the publicly available depth datasets for human action recognition were introduced in detail.(2) The traditional actions representation method using motion history image was modified and a global human actions representation method called multi-layered motion history images (MLMHIs) based on depth information was realized. The 91 transform was then employed to extract feature vector from every layered-MHI and SVM classifier was used for recognition. Experimental results on PMD dataset established by the thesis and MSR Daily3D dataset demonstrated the effectiveness of the method.(3) A local human actions representation method based on spatio-temporal interest points (STIPs) was researched. The STIPs detection methods based on Harris detector and Gabor filter detector were applied on depth information. Then the STIPs were described by cuboid descriptor and 3DSIFT was employed to extract features. Based on STIPs of depth information, the actions recognition was realized.