An Action Recognition Method Based on the Fusion of Neighborhood Gaussian Structure and Video Features

Abstract

The invention discloses an action recognition method for fusion of neighborhood Gaussian structure and video features. In this method, 3D LARK operator is used to extract the local structure features of the video. In order to express the overall structure, a neighborhood structure evaluation algorithm based on multidimensional Gaussian fitting is proposed. Secondly, the neighborhood Gaussian structure and 3D LARK features undergo local matching and statistical processes of the multi-scale template and the video to be tested, respectively, to obtain the statistical probability matrix of the existence of two target actions. Finally, two statistical probability matrices are fused to extract objects, and the double constraints improve the accuracy of object action existence. The present invention proposes the idea of ​​neighborhood relation constraining the whole on the traditional LARK operator, and proposes a new action recognition model. Compared with the existing method, the target action extracted by the present invention is more accurate, the recognition accuracy is higher, and it is suitable for visible light and infrared video of various complex scenes.

Description

technical field [0001] The invention belongs to the moving target recognition technology in the field of computer vision, in particular to an action recognition method based on fusion of neighborhood Gaussian structure and video features. Background technique [0002] Improving the accuracy of target recognition in video is the relentless pursuit of image science research. Efficient computer automatic target recognition technology is of great significance to public security and other fields. The process of target recognition is mainly divided into two methods: training and non-training. The recognition of traditional training methods is heavily dependent on the number of samples, and the classification process is prone to over-fitting problems. At this stage, target recognition technology mainly adopts new non-training methods. [0003] The LARK feature was proposed by Seo et al. in 2010. Compared with HOG features, LBP features, Haar features, SIFT features, etc., it has r...

AI Technical Summary

Problems solved by technology

However, Seo's method uses a single-scale template and cannot identify multi-scale targets

The template contains the background, and the overall matching of the target and the template is used, which results in limited applicable scenarios of the video to be tested, and the detection effect is not good for videos that are not similar to the background of the template.

Images