Personal safety-based human body behavior identification method for infrared video

Abstract

The invention relates to a personal safety-based human body behavior identification method for an infrared video. The method comprises the steps of firstly from night vision human body behavior videosused for training, obtaining static image data streams and dynamic optical flow data streams, and behavior classification tags corresponding to the videos; secondly inputting static gray images, dynamic optical flow images and corresponding tags to a space convolution neural network, and inputting the dynamic optical flow images and the corresponding tags to a time convolutional neural network for carrying out iterative learning, thereby obtaining model parameters of a space network and a time network; thirdly inputting the gray images used for testing and the optical flow images to a space network model for obtaining a result I, and inputting the optical flow images to a time network model for obtaining a result II; and finally performing weighted summation on the result I and the resultII to obtain a final video classification result. According to the method, human body behavior actions in the infrared video can be accurately identified.

Description

technical field [0001] The invention relates to the technical field of infrared image processing, in particular to an infrared video human behavior recognition method based on personal safety. Background technique [0002] Video human behavior recognition refers to the need for the computer to determine the category of human behavior in the video. The ultimate goal of behavior recognition is to analyze who is in the video, when, where, and what they are doing. Correctly judging the category of human behavior is an important step in the further development of behavior recognition. [0003] Video human behavior recognition technology is mainly divided into two categories. The first type is the traditional manual extraction of two-dimensional or three-dimensional features of video images, such as HOG, SIFT, HOF or HOG3D, SIFT3D, IDT and other features, and then classify behaviors and actions through SVM, random tree and other classification methods; the second type is A neura...

AI Technical Summary

Problems solved by technology

In the field of night vision video, Shao Yanhua et al., aiming at the problem that a single feature can express limited human behavior representation information, combined the histogram of orientation gradient (HOG) feature, histogram of optical flow (HOF) feature and motion The Boundary Descriptor (MBH) feature has achieved good results on the infrared data set on 10 categories through the K-nearest neighbor classifier, but the method is too complicated

Images