Research On Deep Learning-based Human Recognition Method For Chemical Production Scenarios

Actions are an important type of information in the real world,which contain people’s emotions and are concrete expressions of their thoughts.Therefore,the research of action recognition is of great significance.Currently,the research results of action recognition have been successfully applied in the fields of human-computer interaction,unmanned driving and virtual reality,and play a key role in chemical production scenarios.By monitoring and analyzing the operations performed by different personnel in the chemical production process in real time,action recognition can function by improving process control,enhancing safety,increasing productivity,and optimizing decision making,and driving intelligence in the chemical production industry.Therefore,action recognition methods for chemical production scenarios have important research significance and practical value.This paper summarizes and concludes existing work on action recognition,including supervised learning and semi-supervised learning-based action recognition,which include related works on both traditional and deep learning-based methods,respectively.In addition,this paper introduces curriculum learning that can effectively reduce the difficulty of model training,which is instructive for the understanding of actions of different complexity.Based on the above work,this paper focuses on stateof-the-art deep learning-based methods for fully supervised action recognition and semi-supervised action recognition,which correspond to the problems of low model recognition ability and less labeled data in chemical production scenarios,respectively.To overcome these problems,this paper proposes a two-stream network-based fusion support vector machine to improve the recognition performance of the model;in addition,semi-supervised action recognition based on curriculum learning and temporal augmentation,and semi-supervised action recognition based on multi-view parallel cooperative learning are proposed to solve the difficulties caused by less labeled data.In this paper,the method design and experimental validation are carried out for the generalization ability of the action recognition model and the low labeled data respectively,which have heuristic implications for accurate action recognition in chemical production scenarios.The main contributions of this paper are as follows:1.This paper proposes a two-stream network fusion support vector machine for action recognition based on the improved two-stream network as a feature extractor,a parallel feature fusion method to obtain hybrid features with stronger characterization ability,and a multi-classification support vector machine with a pair of remaining strategies to replace the feed forward layer in the network for training and classification,and the experimental results verify the effectiveness of the method.2.This paper proposes a semi-supervised action recognition method based on curriculum learning and temporal augmentation,designs a video-based data augmentation strategy to achieve different granularity representations of the same action,constrains the feature representation of the action based on the consistency regularization strategy to achieve the tautological understanding of the model,and sets dynamic thresholds for different categories of actions to reduce the training difficulty of the model in a planned manner,and the experimental results validate the effectiveness of the method.3.This paper proposes a semi-supervised action recognition method based on multi-view parallel collaborative learning,combining the created historical temporal gradient view with RGB view to achieve information complementarity among different modalities,forming a multi-functional committee to evaluate the confidence and consistency of unlabeled data,and designing the average regularized deactivation to reduce the inconsistency of training and prediction caused by deactivation,and the experimental results verify the method The experimental results verify the validity of the method.4.To address the realistic problems of multiple perspectives,complex actions and unbalanced samples in chemical production scenarios,five standard public datasets,including two small datasets,two medium-sized datasets and one large dataset,are selected in this paper.The above datasets contain a large number of potential actions in chemical scenes,and the method in this paper still has comparability with advanced performance methods on the above datasets.