Human behavior identification method based on three-dimensional convolutional neural network and transfer learning model

Abstract

The present invention relates to a human behavior identification method based on a three-dimensional convolutional neural network and a transfer learning model. The method comprises: performing sampling frame by frame of video, stacking obtained continuous single-frame images on a time dimension to form an image cube with a certain size, and taking the image cube as the input of the three-dimensional convolutional neural network; performing training of a basic multi-classification three-dimensional convolutional neural network model when implementation is performed, selecting part of classes of input samples from a test result to construct a sub-data set, training a plurality of dichotomy models on the basis of the sub-data set, and selecting a plurality of models with the best dichotomy result; and finally, employing the transfer learning to transfer the knowledge learned by the models to an original multi-classification model to perform retraining of the transferred multi-classification model. Therefore, the multi-classification identification accuracy is improved and human behavior identification with high accuracy is realized.

Description

technical field [0001] The invention relates to image processing technology in video, in particular to a human behavior recognition method based on a three-dimensional convolutional neural network and a migration learning model. Background technique [0002] In today's society, with the rapid development of storage devices, Internet technology and social networks, large-scale video data has been generated. How to use these video data for target recognition and behavior analysis has become a growing demand. Whether it is intelligent security monitoring, customer shopping behavior analysis, smart home systems, somatosensory games, or the action recognition of pedestrians on the road during unmanned driving, they all rely on high-efficiency and high-precision human behavior recognition systems. The purpose of human behavior recognition is to classify and recognize the behavior or action of one or more people in the video. The research object is often a series of video sequence...

AI Technical Summary

Problems solved by technology

Considering the distribution of behavior in time and space dimensions, different individuals may have different performances for the same type or type of action. These differences are often caused by different individuals' personal understanding of the same type or type of action. Therefore, there are often large differences in the integrity, motion trajectory, motion range, and motion speed of the same action, resulting in difficulties in distinguishing; (2) There are often many interference information in dynamic video information, and many of these information are in the process of static image processing. did not appear in

For example, occlusion between individuals, between individuals and the background environment, relative motion between individuals or between individuals and the background environment during the occlusion process, changes in light intensity and contrast in the video sequence, video The movement and zoom of the lens during shooting, and the action in a video sequence may be incomplete, etc.

All these difficulties have brought great challenges to the research of behavior recognition in video sequences.

Images