A method and system for human action recognition in video

Abstract

The invention discloses a method and a system for recognizing human behavior in video. An optical flow frame generation network and a long-term recursive convolutional neural network are cascaded to form a time-flow deep convolutional neural network, and a long-term recursive convolutional neural network constitutes a space. The streaming deep convolutional neural network uses a multi-dimensional weighted fusion model to fuse the prediction results of the dual-stream network to obtain the prediction of the human behavior type of the video data. The invention has high detection accuracy, wide application occasions and good generalization ability.

Description

technical field [0001] The invention belongs to the field of computer vision and machine learning, and particularly relates to a video recognition method. Background technique [0002] Human behavior recognition in video data has important theoretical research value and potential application value in the fields of intelligent video surveillance, smart home, human-computer interaction systems, and content-based video retrieval. Taking intelligent video surveillance as an example, although video surveillance systems have been popularized in important occasions such as transportation, power systems, and buildings, they have not achieved true intelligence, that is, computers can autonomously understand human behavior in videos, and when abnormal behavior occurs , giving timely and accurate warnings. [0003] Traditional human behavior recognition algorithms are based on manual feature extraction and shallow machine learning algorithms. However, these algorithms can only achieve...

AI Technical Summary

Problems solved by technology

These optical flow calculation methods not only require offline calculation and high computational complexity, but also usually cannot obtain significant optical flow information when the moving target only has a small displacement, which is not conducive to the identification of behavior types

At the same time, the offline calculation method cannot achieve joint optimization with the dual-stream architecture deep neural network, which seriously affects the recognition speed and performance of the algorithm

Images