Human body motion classification method based on compression perception

Abstract

The invention relates to a human body motion classification method based on compression perception, comprising the four steps of space-time interest point detection, video characteristic expression based on a bag-of-word model, construction of a visual dictionary and a motion classification algorithm based on compression perception. In step 1, solving training sample characteristics to obtain a training sample matrix A=[A1,A2,...AK] belong to Rm*n, k categories, a test sample y belong to RM and an optional fault tolerance degree epsilon>0; in step 2, solving a dictionary Z, a classifier W and a coefficient matrix A; and for a new video motion sequence, employing the classifier W obtained in the second step for classification, and finally obtaining the category estimation of video motion. The human body motion classification method fuses space-time interest detection, dictionary learning and video expression characteristics in a learning framework, and simultaneously learns a linear classifier; the human body motion classification method simultaneously learns a discrimination dictionary, a discrimination coding coefficient and a classifier through an optimal method, is simple to calculate, has good robustness, and enhances the capability of processing non-linear data through a compression perception method.

Description

technical field [0001] The invention relates to a method for classifying human actions, in particular to a method for classifying human actions based on compressed sensing, and belongs to the field of video analysis. Background technique [0002] It is well known that extracting data from videos to make reasonable representations of actions is especially important for action classification. Usually we need to select the method of action representation according to the method of action classification. For example, trajectory-based methods are suitable for long-distance monitoring in open environments, while 3D models are often used in gesture recognition. Parameswaran et al. proposed to use the following four criteria to evaluate action representation methods: simplicity, completeness, continuity, and uniqueness. [0003] The shape of the human body contour is the most intuitive way to represent actions, so there are also a large number of shape-based human action represent...

AI Technical Summary

Problems solved by technology

Jiang et al. used the features generated by the Action bank detector as the feature representation of the video, which relied on pre-trained detectors, and the accuracy was not high.

Images