Automatic Face And Action Recognition In Construction Site Monitoring Video

In order to adapt to the increasing engineering application of intelligent video surveillance,research in areas such as automatic detection and alarm of video in complex scenes is booming.This paper is based on the intelligent video surveillance project,taking the construction site as the research scenario,and mainly investigates and researches on a series of methods related to face detection,face recognition and human behavior recognition in surveillance video.The transformation of the theoretical results related to artificial intelligence into actual intelligent video surveillance products studied in this article has positive significance for both academic research and practical applications.This paper focuses on face detection,face recognition and human behavior recognition in construction site surveillance videos.It mainly researches on the face detection technology that uses multi-face targets and small face targets as detection targets.And it also researches on face recognition technology and human behavior recognition technology.The main research content is divided into four parts.First of all,according to the problems of small and many face targets and different target scales in the construction site scene,a face detection algorithm based on single-stage neural network is introduced.Through the scale-invariant design and context module design in the algorithm,the loss function design in training and the online hard case mining,the algorithm can solve the target multi-scale problem and the small target and multi-target problem in this task.This paper proves the superiority of the algorithm in accuracy through experiments on public data sets and the feasibility of accuracy and speed in project application.Secondly,according to the detected face as a recognition object,the problem of insufficient resolution of the face,excessive noise interference resulting in insufficient face information,introduces and improves the face recognition algorithm based on additional angle boundary loss.By adding an angular margin loss during the network training process,the network is easier to converge and has better robustness.It can also make the distance between the same classes smaller and the distance between different classes larger,and improve feature discrimination.The algorithm can solve the problem of insufficient face information mentioned above,and extract the separable face features from the object to be tested.The experiments on public data sets in this paper prove the superiority of the accuracy of the algorithm and show it in the project application.Feasibility of accuracy and speed.In addition,in order to make the task more robust in the project,this paper also designed a multi-frame face recognition process.Finally,according to the problems of difficult to recognize human behavior in the actual scene and many complex interference factors in the scene,the spatial temporal graph convolutional networks for skeleton based action recognition was introduced.By extending the concept of convolutional neural network from traditional image convolution to three-dimensional convolution in time and space domain,in addition,the method of graph convolution and subset division strategy are proposed,so that human action recognition for bones can be better through neural network automatically extracts behavioral features.The superiority of the algorithm and the feasibility in the project have been proved through experiments on public data sets.In order to solve the interference problem of complex scenes,this paper also designs a secondary action recognition process.