Research On Theory And Algorithm Of Human Action Recognition Based On Video

Human action recognition based on video is an active research branch of computer vision technology.It has a wide application prospect in the fields of intelligent surveillance,behavior analysis,video retrieval and human-computer interaction.The application of action recognition is limited to gesture recognition and some simple body action recognition at present.And the large-scale application maturely still has a long way to go.This thesis mainly studies human action recognition from two aspects: traditional machine learning and deep learning.The main content of this thesis is organized as follows:1.A novel feature descriptors fusion method based on dense trajectory algorithm is proposed.Dense trajectory can capture complex motion information effectively,and the algorithm based on dense trajectory has been proven to be very effective in action recognition.We introduce 3-Dimension Histogram of Oriented Gradient(3DHOG)feature descriptor in the framework of dense trajectory algorithm,which is the three-dimensional expansion of Histogram of Oriented Gradient(HOG)feature on the spatial-temporal cube.And we fuse 3DHOG with Histogram of Oriented Flow(HOF)and Motion Boundary Histograms(MBH).The fusion improves the accuracy compared with the original combination of ‘HOG+HOF+ MBH'.2.Low-rank sparse coding based on Fisher regularization-constraint.Low-rank sparse coding is an effective dictionary coding method,which has sparsity,locality and spatial consistency characteristics.We introduce the Fisher regularization-constraint on the basis of sparse low rank coding,which makes the coding has the advantage of low-rank sparse coding and keep the properties that is smaller within-class scatter but bigger between-class scatter,so that enhance furtherly the discriminating ability of the coding vector.3.In order to overcome the problem of poor accuracy of deep network model implemented on lightweight human action recognition data set,a novel action descriptor based on combination of 3D convolution deep feature and hand-craft feature is proposed.4.3D convolution neural network is combined with two-stream convolution neural network.3D convolution neural network and two-stream convolution neural network are classical deep learning models in the field of action recognition.The combination of them can improve the effectiveness.