Single frame representation model based human activity identification method

Abstract

The invention provides a single frame representation model based human activity identification method. The single frame representation model based human activity identification method mainly comprisespretreatment, a single frame representation model, an activity identification model, and model optimization and training; the process is as follows: generating a light stream image for each video frame of an inputted video first, then inputting all videos and the corresponding light stream image to one single frame representation model, generating model representation, then using a long and shorttime memory model to generate a predicated final activity label according to the model representation generated in the previous step, and finally determining the final activity label through a full connection layer with a Softmax activation function. The single frame representation model based human activity identification method solves the problems that the previous identification precision of the human activity is not high, the identification requires long computation time and the group activity cannot be accurately identified, the method can identify group activities, the identification precision is relatively high, and the required computation time is less.

Description

technical field [0001] The invention relates to the field of activity recognition, in particular to a human activity recognition method based on a single-frame representation model. Background technique [0002] With the advancement of embedded system design technology, powerful cameras have been embedded in various smart devices, and wireless cameras can be easily deployed on street corners, traffic lights, large stadiums, railway stations, etc., resulting in a large number of Videos have attracted a large number of researchers to study human activity recognition. Human activity recognition can be applied to law enforcement tasks responsible for monitoring large-scale crowd activities. Through rapid human activity recognition and analysis on a large number of videos captured by street surveillance, suspicious or criminal behavior can be quickly identified; similarly, if TV broadcasts can automatically identify and broadcast games Highlights of the video, rather than the en...

AI Technical Summary

Problems solved by technology

[0004] In view of the low accuracy of human activity recognition in the past, the long calculation time required for recognition, and the inability to accurately identify group activities, the purpose of the present invention is to provide a human activity recognition method based on a single-frame representation model. Each video frame of the video generates an optical flow image. Then, all video frames and corresponding optical flow images are input into a single frame representation model to generate a model representation. Then, a long short-term memory model is used to generate according to the previous step. The model representation generates the predicted final activity label, and finally, a fully connected layer with Softmax activation function is used to determine the final activity label

Images