Human Action Recognition Based On Dense Trajectory And Regularized Multi-task Learning

In latest researching of human action recognition,an increasing number of researchers have paid close attention to the aspects of excavating related information of multi-view?multi-modal and different action.The reason of this situation was caused is human action recognition of RGB was easily influenced by some factors in traditional single view.These factors consist of illumination condition,view condition and the differentiation of human action and so on and restrict development and application of researching work further.Therefore,the paper is composed of three parts as follow: 1)proposing human action description algorithm of deep motion trajectory information in single view;2)evaluating and analyzing that different regularized multi-task learning algorithms exert an effect on human action recognition;3)coming up with multi-view human action recognition algorithm based on image set and regularized multi-task leaning.Concrete details as follow:1)Human action description algorithm based deep motion trajectory information.Firstly,the algorithm excavates dense trajectory feature from deep video by means of the various information of optical flow,adding deep information;Secondly,segmenting the cube of dense trajectory into several cubes and computing whose descriptors-HOG(Histograms of Oriented Gradients),HOF(Histograms of Optical Flow)and MBH(motion boundary histogram)that can eliminate disturbance camera motion caused were computed;Finally,K-means,BOW(Bag-of-Word)and SVM(support vector machine)model was exploited to make projection and classify.The description algorithm has more favorable experiment performance by the result of public deep action dataset-DHA-17 and UTkinect.2)Evaluating and analyzing performance of regularized multi-task leaning algorithm in the field of human action recognition.As the relatedness of different tasks can be excavated by using regularized multi-task leaning algorithm,various actions were considered as different tasks in different view therefore.Utilizing several the functions of regularized multi-task leaning algorithm to dig related information each other.These results compared with the single task learning algorithm.In addition,the performance of these functions was further evaluated in cross-view task.The performance of three datasetsCVS-MV-RGBD-SINGLE,IXMAX and UCLA – has proved that regularized multi-task leaning algorithm have superiority and robustness in the aspect of excavating related information of different action in single view a certain extend.However,the conclusion is opposite in cross-view,whose reason is the otherness exists between different view.3)The algorithm of multi-view human action recognition model based on image set and regularized multi-task learning.In order to verify that multi-view sample exerts an effect on regularized multi-task learning model,the algorithm excavates relatedness of different view and action by adding training samples of different view.Subsequently,image set algorithm and multi-task learning function(Least_RMTL)were used simultaneously to dig related information further.The performance of three datasets-CVS-MV-RGBD-SINGLE,IXMAX and UCLA – has proved that the increase of sample number of multi-view can enhance performance of multi-task learning algorithm.Simultaneously,the fusion of image set algorithm and multi-task learning model can excavate relatedness of different action in multi-view further.