Behavior identification method for rank-1 tensor projection based on canonical return

Abstract

The invention discloses a behavior identification method for a rank-1 tensor projection based on canonical return. The method disclosed by the invention comprises the following steps: (1) the training course: two two-class classifiers are respectively trained for each behavior by adopting a One Vs All strategy; a projection vector set comprising three projection vectors is solved by utilizing a canonical return method; the rank-1 tensor projection is obtained through the exterior product of the projection vector set; the training behavior with the known class is divided into multiple subsets according to the types; the embedding of the subsets is established; and (2) the identification course: testing data are input to each two-class classifier; the output responses of all the two-class classifiers form a row vector; the row vector is multiplied by the embedding result of each one-dimensional subset to obtain the maximum row vector of the multiplying result; and the positive sample class represented by the two-class classifiers corresponding to the maximum row vector is the class judged by a testing behavior. In the method disclosed by the invention, the behavior identification rate is high; and the computing complexity of the method can be greatly reduced.

Description

technical field [0001] The invention relates to the fields of computer vision and image processing, in particular to a behavior recognition method based on rank-1 tensor projection of regular regression. Background technique [0002] Human behavior recognition is an important research direction in the field of computer vision. Its purpose is to enable computers to understand human behavior like humans. At present, this technology is widely used in intelligent monitoring, virtual reality, human-computer interaction and motion analysis, etc., but when the scene is complex, the result of human behavior recognition is far from people's expected goals The needs of practical technology. [0003] Human behavior exists in video sequences, which are three-dimensional data containing spatial and temporal information. The traditional method of processing video data is usually to convert each frame of image in the video into a single vector, and then concatenate it into a matrix in th...

AI Technical Summary

Problems solved by technology

But it loses timing information and is sensitive to different execution rates between behaviors

[0007] (4) Kellokumpu et al. (V.Kellokumpu, G.Zhao and M.Pietikainen, Human activity recognition using a dynamic texture based method, BMVC2008) used the low-dimensional texture features of the local binary pattern extraction behavior on three orthogonal planes to perform recognition, but the background, actor appearance, clothing, etc. in the original behavior data will cause messy texture information, and the space-time information on the three orthogonal planes is not enough to describe the effective characteristics of the entire human behavior

[0008] There are two main defects in the above method: (1) Zhang Quantum Space method uses High Order Singular Value Decomposition (High Order-SVD) method to solve the optimization problem, resulting in huge time consumption

(2) If the compact representation of the original video data is called "feature tensor", its general form is also three-dimensional. For the feature tensor of three-dimensional volume, which measurement method is the best has not been verified

[0009] The above two major defects make the existing behavior recognition method represented by tensors fail to meet the practical requirements of behavior recognition in terms of recognition performance and time efficiency, which limits the scope of application of this method.

Images