Behavior Recognition Method Based on Joint Statistical Descriptor in Wavelet Domain

Abstract

The invention discloses a behavior recognition method based on joint statistical descriptors in wavelet domain, which mainly solves the problems of insufficient time-space information extraction, insufficient consideration of relationship between coefficients and low robustness in existing wavelet domain behavior recognition technology. The technical solution is: 1. Extract the dense trajectory of the behavioral video, and construct a cube curved along the trajectory; 2. Decompose the video with 3D stationary wavelet transform to obtain coefficient subbands in each direction; 3. In the cube curved along the trajectory , to construct the mutual information descriptor between the wavelet coefficient subbands; 4. In the cube curved along the trajectory, construct the co-occurrence histogram descriptor of wavelet coefficients at all levels according to different neighborhood directions; 5. Concatenate the mutual information descriptor and the co-occurrence histogram The histogram descriptor is obtained, and the joint statistical descriptor in the wavelet domain is obtained; 6. The bag-of-words model of the descriptor is constructed, and the SVM classifier is used for identification. The invention improves the accuracy rate of human behavior recognition and can be applied to intelligent monitoring and human-computer interaction.

Description

technical field [0001] The invention belongs to the technical field of video processing, and further relates to a behavior recognition method, which can be used for intelligent video monitoring and human-computer interaction. Background technique [0002] In the previous decades, human behavior recognition and its related research fields have always been considered to be very challenging scientific research directions in the computer vision discipline. Human behavior recognition allows computers to learn, analyze, understand and memorize human behavior through algorithm design, and realize computer classification and discrimination of human behavior videos. Technologies related to human behavior recognition are widely used, such as camera monitoring, multimedia semantic annotation and indexing, pedestrian tracking and human-computer interaction, etc. More and more scholars and institutions have carried out a lot of research work in related fields. The essence of behavior cl...

AI Technical Summary

Problems solved by technology

This method first extracts spatio-temporal interest points, and then performs two-dimensional wavelet decomposition in the cube around the interest points to obtain discriminative and reliable descriptors. The feature dimension is low, and it has certain tolerance to the influence of noise and illumination. However, this method only selects three representative planes in the local cube when performing wavelet decomposition, which does not cover enough data, and using the method of interest point extraction is a direct extension of two-dimensional analysis to three-dimensional, ignoring the time domain structure and space domain structure difference, the obtained space-time information is insufficient

In addition, the method only connects the wavelet coefficients in series when constructing features, without considering the characteristics of the coefficient direction and the relationship between coefficients.

Images