Research On Behavior Recognition Of Factory Personnel Based On Machine Vision

With the rapid development of the industrial era,there is an increasing demand for factory production efficiency and safety.To evaluate the efficiency of various production processes,and to avoid accidental injury accidents,video monitoring,and intelligent analysis technology have played a very important role.However,the existing traditional monitoring systems still have some shortcomings.Firstly,traditional video monitoring mostly uses manual methods to judge the behavior of personnel in various production links,which is not accurate enough;Moreover,traditional video monitoring technologies are mostly offline,unable to accurately and real-time judge personnel behavior and risk early warning.This article aims at these two main issues,starting from three aspects,to conduct research and analysis on personnel behavior during factory production.Firstly,an identity recognition algorithm based on metric learning is proposed.The algorithm uses a residual network to construct a data prediction model,extracts the person’s feature information from the image,and then compares the features through Euclidean distance calculation to identify the person’s identity.The high point of this algorithm is that it can improve the recognition accuracy and enhance the customization ability of the system from two parts: feature extraction and Euclidean distance determination.Compared to the general idea of using multiple target tracking algorithms to identify a person’s identity,this algorithm can also track across cameras,avoiding the problems of identity loss and identity jump to a certain extent.Finally,the practical feasibility of the proposed identification algorithm was verified under the Market1501 dataset and the GENER-Market dataset,respectively.Secondly,to address the issue of personnel behavior within the factory,a SlowFast behavior recognition algorithm that integrates the Convolutional Block Attention Module(CBAM)has been proposed.The algorithm uses two parallel convolutional neural networks to process the same video segment,extract spatial and behavioral feature information respectively,and then fuse features through lateral connections to identify human behavior.Compared to traditional behavior recognition algorithms,this mechanism effectively reduces the training and reasoning time of the network.In order to further improve accuracy,the method of knowledge distillation is also used to improve the network by replacing SlowFast slow channels with motion simulation streams.Through comparative experiments on user-defined datasets and UCF-101 datasets,it is found that the improved SlowFast network has lower overall complexity and higher prediction accuracy compared to other methods.In addition,the behavior of personnel in a factory often requires multiple factors,so it is not sufficient to analyze the behavior of only one person.To address this issue,this article also designed an interactive software that integrates identity detection and behavior recognition.The software sets norms for multi person behavior,establishes a multi-person behavior recognition system,and transforms deep learning tasks into hierarchical analysis tasks,taking into account changes in scene factors.The system is designed with a "decentralized identification and centralized analysis" structure.Compared with traditional behavior recognition algorithms that are completely based on deep learning,the system has a certain degree of flexibility and better detection results.