Research On Dangerous Behaviour Recognition Technologhy Based On Monitoring Areas

To protect people’s property and personal safety,video surveillance equipment is used in large numbers in various public places and other key areas.Although the use of surveillance systems can help relevant personnel to carry out business supervision,effectively eliminate the occurrence of risks,play the role of a case barrier,and also play an important role in evidence as evidence in relevant cases.However,human energy and subjectivity cannot guarantee a rapid and accurate judgement of relevant abnormal behaviour under a large amount of video data for a long time,and even in the case of accurate identification,both sides of the incident can still be harmed if not stopped in time.With the development of computer vision and deep learning technology,human behaviour recognition for surveillance video for information extraction has been widely studied and achieved good recognition results.However,due to the influence of external factors such as complex and diverse backgrounds in the video,the network model usually suffers from difficulties in accurate recognition,large number of parameters and redundant feature information extraction.Therefore,this paper makes a discussion on how to use surveillance equipment for all-round monitoring.The surveillance system mainly includes the prediction function for the interactive objects beforehand and the study of the behaviour recognition algorithm for the behavior during and afterwards,the main contents are as follows:(1)To address the problems of redundant information extraction and insufficient algorithm accuracy in feature extraction in the gaze follow network,this paper adopts the principle of changing the feature extraction method by replacing the convolutional neural network used in the original network with the Swin-Transformer.The improved model was trained on The Gaze Follow dataset and Video Attention Target dataset,and the detection accuracy of the whole model reached 92.3% and 89.3% respectively after training on the dataset,and the results were very close to the real accuracy of humans themselves.Based on the gaze follow model,this paper also proposes a pedestrian interaction object prediction model based gaze follow technology,which learns the level of attention to the interaction object by combining the statistics of each frame of the video sequence with the training network model to learn the level of attention to the interaction object in the data before the interaction occurs with the interaction object,and combines the results of both to obtain the final interaction object prediction.(2)A human behaviour recognition algorithm based on 3D-Ghost module with global axial self-attention mechanism is proposed for the problems of large number of parameters of 3D convolutional neural network,redundancy of feature information extraction and improvement of accuracy.In order to reduce the number of parameters in the model,this paper extends Ghostnet from a two-dimensional convolutional neural network to a three-dimensional convolutional neural network and replaces the convolutional kernel contained in the Inc.module of the original I3 D network,while adding a residual channel to optimize the network structure to obtain the final lightweight module 3D-Ghost module and apply it to the whole network.A global axial attention module based on the spatio-temporal separation self-attention mechanism is incorporated into the improved I3 D network to improve the behavioural recognition accuracy of the model.Experimental results show that the proposed model network has a computational volume of 14.85 GFlops and a number of parameters of 18.83 M.The recognition accuracy on UCF101 is 96.2%,which proves that the algorithm model effectively reduces the number of parameters and improves the accuracy.(3)Building an environmental monitoring system for public places.By experimentally testing the interactive object prediction system and the behaviour recognition system in real public area places,an environmental monitoring system is designed in this paper.The system can make a choice of models and video’s and call different detection algorithms.The camera can also be used for real time monitoring functions.The system has been tested on the basis of home-made real public surveillance video data and the results of the system also prove the feasibility of the algorithms in this paper.