Human action recognition method based on double layers of conditional random fields

Abstract

The invention discloses a human action recognition method based on double layers of conditional random fields, and belongs to the field of action recognition of computer vision. Firstly, a human posture of an action main body in a RGB-D video and information features of objects possibly interacting with the action main body are respectively extracted, and score information of each small video of the interacting objects, which is obtained after the RGB-D video is split, is calculated and used as a global feature. Then, a top layer conditional random field is modeled to capture high-order dependence among human actions, a bottom layer conditional random field is modeled to enrich potential structures inside the human actions, and finally, a discrimination and classification of double layers of conditional random fields is constructed. Next, precise inference and a structured support vector machine classifier are adopted to learn parameters of the discrimination and classification of double layers of conditional random fields. Finally, according to the model parameters obtained by learning and a vested model, a human action category in a tested video is predicted. According to the invention, recognition accuracy of the human actions is improved to a certain degree.

Description

technical field [0001] The invention relates to the technical field of computer vision action recognition, in particular to a human action recognition method based on double-layer conditional random fields (Double-layer conditional random fields model for human action recognition, DL-CRFs). Background technique [0002] Human behavior recognition in video sequences is a research topic involving computer vision, pattern recognition and artificial intelligence. It has been a hot research topic because of its wide application value in business, medical and sports fields. [0003] Literature [Koppula H S, Gupta R, Saxena A. Learning Human Activities and Object Affordances from RGB-D Videos [J]. International Journal of Robotics Research, 2013, 32 (8): 951-970.] according to the complexity of human behavior Behaviors are divided into high-level activities and simple actions. Simple behavior refers to an indivisible behavior with at most one interactive object in the process, and...

AI Technical Summary

Problems solved by technology

The obvious shortcoming of the generative model is that when there are complex correlations between the input sample data, the modeling of the joint distribution will become complicated or even inaccurate

However, the above-mentioned existing behavior recognition methods based on probabilistic graphical models have not considered both the internal potential structure of the human behavior state and the high-order correlation between human behavior states, and there is still the problem of low recognition accuracy.

Images