Method for detecting video action based on convolutional neural network

Abstract

The invention publishes a method for detecting video action, and relates to the technical field of computer vision recognition. The video action detection method is based on a convolutional neural network, a space-time pyramid pooling layer is added to a network structure, restrictions of a network on input is eliminated, the speed of training and detection is improved, and the performance of video action classification and time positioning is improved; the convolutional neural network includes convolutional layers, common pooling layers, space-time pyramid pooling layers and full connection layers; and output of the convolutional neural network includes a category classification output layer and a time positioning calculation result output layer. According to the method provided by the invention, video clips of different time lengths do not need to be obtained through downsampling, but by direct one-time input of a whole video, thereby improving efficiency; and at the same time, since the network trains video clips of the same frequency, difference within a category is not increased, the learning burden of the network is reduced, model convergence is relatively fast, and a detection effect is relatively good.

Description

technical field [0001] The invention relates to computer vision recognition technology, in particular to a video action detection method based on a convolutional neural network. Background technique [0002] In recent years, almost everyone has a mobile phone. Every mobile phone has a camera. Coupled with the development of the Internet and the advancement of communication technology, people are more and more fond of shooting videos and spreading them on the Internet, so the number of videos is growing explosively, and video storage and analysis technology is very important. [0003] Video motion detection refers to classifying the motion in the video and giving the start time and end time of the motion, for example figure 1 shown. In recent years, the task of video action recognition has made great progress, but it is mainly suitable for cropped videos, that is, there is only one action in the video, and there are no redundant frames. As a result, scholars began to stud...

AI Technical Summary

Problems solved by technology

[0011] (1) Sparse sampling will destroy the continuity within the action, so that the 3D CNN cannot learn better motion features;

[0012] (2) The video clips obtained at different sampling frequencies share a network for training, which will increase the intra-class differences, increase the network learning burden, and require more complex networks and more training data.

Images