A Feature Coding Method for Spatio-temporal Interest Points in Human Action Recognition

Abstract

The invention discloses a feature encoding method of spatio-temporal interest points in human action recognition. Human action recognition in video has broad application prospects in intelligent monitoring and video retrieval. The human action recognition method based on spatio-temporal interest point features has been widely concerned and valued because of its simple feature extraction, strong anti-interference ability, and good robustness. However, in this method, in the process of encoding spatio-temporal interest point features by vector quantization to obtain video representation vectors, there are problems such as large representation error, weak encoding discrimination ability, and loss of spatial position information. In order to solve these problems, the present invention proposes a space-regularized local constraint coding algorithm. In the process of feature encoding, local constraints are introduced to reduce representation errors, and spatial regularization is introduced to enhance the differentiation of encoding results and use the feature space position information of spatiotemporal interest points to finally improve the accuracy of human action recognition.

Description

technical field [0001] The invention mainly relates to the field of digital video content understanding and analysis, in particular to a feature encoding method of spatio-temporal interest points in human action recognition. Background technique [0002] Human action recognition in video is to extract visual information that can describe the characteristics of human action actions from video sequences, and use machine learning algorithms to classify these information to achieve the purpose of identifying human action actions. It has broad application prospects in intelligent monitoring, video retrieval, robot control, etc. [0003] Human action recognition in video includes video feature extraction, action representation, action classification and other links. The video features used for action recognition mainly include: human body model, global features and local features. Compared with human body model and global features, local features have the advantages of simple ex...

AI Technical Summary

Problems solved by technology

Second, the distribution of spatio-temporal interest point features in its representation space behaves as a nonlinear manifold (3), and vector quantization is difficult to model the nonlinear manifold (3) structure

However, vector quantization only utilizes their feature description information, while ignoring their location information in the human body area, thus, it cannot model the spatial location distribution of spatio-temporal interest point features

Among them, the first two shortcomings can be summarized as the problem of large representation error, which is manifested as large quantization error and poor accuracy of nonlinear manifold representation.

The third deficiency can be considered as the encoding fuzzy problem, which is manifested as giving the same encoding result to the spatio-temporal interest point features from different human body parts

The fourth is the problem of missing spatial position relationship, which is manifested as the inability to model the spatial position relationship of spatio-temporal interest point features

Images