Action Representation And Recognition Based On Depth Information

In recent years,computer vision technology has developed rapidly.As an important research subject of computer vision,behavior recognition has important scientific significance.As we all know,the action recognition of visible image is often affected by interferences,such as illumination,shadow and occlusion.There is no obvious change in two-dimensional plane for some action.However,it has great change in depth direction actually.Hence,the action will not be recognized exactly just according to the RGB image.In addition,most of the existent method based on depth information use the nonzero pixels to calculate depth difference,and some are only according to the change of depth direction.A method of action description based on point cloud-Motion History Point Cloud(MHPC)is proposed.It is similar to Motion History Image(MHI)action sequence compression as a two-dimensional image;MHPC compresses action sequence into three-dimensional point cloud which includes the depth information.The channel(x,y,z)of point cloud is used to record the change of the space position of the prospect point in the process of the action,and the color channel is used to record the variation of the time sequence.MHPC has the following two characteristics: first,the data of MHPC is compressed.And the temporal and spatial variation information in the process of action is retained in MHPC.Due to containing the variation of the depth information,the action is depicted more meticulous.So,more rich features can be extracted to support the subsequent action recognition.Second,for the subsequent process of feature extraction and action recognition,MHPC is more convenient than the original video data in processing.Hence,MHPC is a kind of standard input and is also a means to strengthen the pretreatment of movement characteristic.As an application example of MHPC,a method for action recognition based on local feature point extraction,bag of words and multiple classification support vector machine(SVM)is provided.Harris3 D detector is used to extract the feature points from the MHPC generated,and Fast Point Feature Histogram(FPFH)is adopted to describe the feature points extracted.Then,the Bag of words is generated by clustering the feature descriptors.Finally,the action classification and recognition is realized by using the multiple classification support vector machine(SVM).Compared with other similar methods,the effectiveness and reasonability of the means is verified by the experiment.