A multi-view action recognition method

Abstract

The invention discloses a multi-view action recognition method, which includes two processes of action training and action recognition. During action training, the classifier is trained by the two-dimensional conditional random field method; the action recognition process includes the following steps: extracting spatiotemporal interest points; calculating feature descriptors; feature descriptor dimensionality reduction; feature descriptor clustering, and obtaining preprocessed files; Feed the preprocessed file to the classifier obtained during training. The present invention makes full use of the spatio-temporal relationship between spatio-temporal interest points, and effectively describes the characteristics of different actions; adopts K-means clustering to gather different actions into different categories, increasing the discrimination of action recognition; by introducing The two-dimensional conditional random field can effectively model the temporal action sequence under a single camera and the spatial action sequence between multiple cameras, making the training model more accurate, thereby realizing the effective recognition of human action.

Description

technical field [0001] The invention belongs to the technical field of computer vision, and in particular relates to a multi-view action recognition method. Background technique [0002] Using video cameras and computers to "see" instead of human eyes, that is, computer vision technology, is beginning to receive more and more attention. This technology can recognize images and videos and perform further processing by taking pictures with cameras and using preset algorithms in the computer. This technology attempts to establish an artificial intelligence system that obtains and processes information from images or videos. [0003] Moreover, with the increasing maturity of video monitoring technology and the popularization of monitoring equipment, the cost of monitoring equipment such as cameras is decreasing day by day, and the acquisition of video information becomes easier and more convenient, and the quality of video information is also getting higher and higher. Based on...

AI Technical Summary

Problems solved by technology

[0012] Because the existing technology has poor behavior recognition effect on subtle differences, is sensitive to changes in motion duration and noise, and adapts to observation sequences with dependencies, and cannot get good results for more complex motions, only one dimensional linear conditional random field, the present invention proposes a multi-view action recognition method

Images