Human Behavior Recognition Method Based on Asymmetric Generalized Gaussian Model

Abstract

The invention discloses a human behavior recognition method based on an asymmetric generalized Gaussian model. Firstly, the training video is preprocessed and the spatio-temporal interest points of the video are detected, and then its video blocks are extracted centering on the interest point and its optical flow is calculated. Information and gradient information, draw the corresponding histogram according to the obtained optical flow information and gradient information, and then use the asymmetric generalized Gaussian model (AGGD) to fit the corresponding histogram, and use the AGGD parameters of the optical flow information and gradient information as The features form the feature matrix of the training videos. For the test video, all the above-mentioned processes are also performed to obtain the feature matrix of the test video. Finally, calculate the Mahalanobis distance between the feature matrix of the training video and the test video, and then identify the behavior of the test video according to the nearest neighbor principle. The method of the invention greatly improves the accuracy rate of video behavior to be recognized.

Description

technical field [0001] The invention relates to a method for human behavior recognition, which belongs to the field of computer vision and machine learning, and specifically relates to a human behavior recognition algorithm. Background technique [0002] In recent years, the research and application of intelligent video surveillance technology has attracted people's attention. As its basic processing part, action recognition is a very active research direction and belongs to the important research content in the field of computer vision. [0003] According to the current research methods, it can be divided into two categories: behavior recognition methods based on global features and local features. Global features usually use information such as edges, optical flow, and silhouettes to describe the entire detected human body area of ​​interest, and are sensitive to noise, viewing angle changes, and partial occlusions. For example, use the mixed Gaussian model to adaptively...

AI Technical Summary

Problems solved by technology

For various behaviors of the human body, the optical flow information of the limbs with inconspicuous movement can be ignored, and the above method needs to calculate the optical flow of the entire human body area, which not only increases the amount of calculation, but also reduces the recognition accuracy

At the same time, for the spatio-temporal features, the spatio-temporal feature descriptors are subjected to PCA dimensionality reduction and then the bag-of-words codebook is constructed, that is, the training data is sampled and then clustered to generate a "dictionary". This method makes the training samples not fully utilized. ; and in order to ensure a certain average recognition rate, even if the dimensionality is reduced, the amount of sample data is still too high, and the clustering speed is relatively slow

In addition, there may be a certain similarity in the feature data of each direction, and clustering all directions together will reduce the descriptiveness of the behavior of different direction features.

Images