BM-CNN model-based dangerous action identification method and system in short video APP

Abstract

The invention provides a method and system for identifying dangerous actions in a short video APP based on a BM-CNN model, and the system comprises an initialization module which obtains a video, carries out the preprocessing and convolutional neural network processing of the video, carries out the automatic calling when the system starts to run, and can be called again through an administrator module; the administrator module is used for processing the data set and performing manual auditing, the module is called by an administrator, and functions are displayed to a user; the user module is used for providing video uploading and result display functions and interacting with the working module to realize action recognition and classification functions; and the working module is used for receiving videos, realizing action recognition and classification and expanding data sets of different classification actions. According to the method, a dangerous behavior data set is established automatically, and different directivity training can be carried out according to different definition requirements of a short video APP on dangerous actions; according to the invention, independent training is carried out on different actions, so that an alarm function for dangerous actions can be realized, and a classification function for non-dangerous actions can also be realized.

Description

technical field [0001] The present invention relates, in particular, to a method and system for identifying dangerous actions in a short video APP based on a BM-CNN model. Background technique [0002] In 2017, Joao Carreira and Andrew Zisserman proposed the behavior recognition dataset kineticsdataset and I3D (Two-Stream Inflated 3D ConvNets) model, which expanded the 2D convolutional network into a 3D convolutional network for the first time, forming a spatiotemporal (3D) network technology. route. In this route, the video recognition architecture is designed by extending an image classification network with a temporal dimension, while preserving the spatial properties. These extensions include extending directly from 2D models such as ResNet or InceptionNet to 3D, adding RNNs on top of 2D CNNs, etc. [0003] In recent years, the data volume of short videos has increased significantly, and many of them contain short videos of dangerous actions, such as parkour and rock c...

AI Technical Summary

Problems solved by technology

Due to the popularization of smart phones and the penetration of short video platforms, and the fact that young people have not yet formed the correct three views and mature judgment ability, these short videos will have a negative impact on young people, causing them to rush to imitate, causing bad results, causing harm to themselves or others. lives in danger

[0004] Based on this, the applicant conducted further investigation and research and found that the root cause is that there are two major problems in the review of short videos: first, most of the current short video platforms are released first and then reviewed, which has caused a certain lag in supervision The second is that most of the review process is manual review, which leads to relatively low accuracy and efficiency

[0009] At the same time, patent document CN110969130A provides a method and system based on YOLOV3 driver’s dangerous behavior recognition, which acquires the driver’s infrared image, determines the face position through the face detection algorithm, and selects the driver’s dangerous behavior to be recognized area according to the face position, so that the prediction result It is more accurate, but the disadvantages are: firstly, the pre-collected sample data is used for training, and the obtained model tends to be fixed, and there is no room for updating; secondly, the patent does not elaborate on the system functions, and its detection method is different from the system Obviously; moreover, its use scene determines that the types of actions it can recognize are limited, and many other dangerous actions are not included

However, this invention does not establish its own dangerous behavior data set, and it cannot conduct different directional training according to the different definition requirements of short video APPs for dangerous actions.

Images