Research And System Implementation On Action Compliance Recognition Based On Deep Learning

As the terminal market of fuel consumption,the refueling area has high requirements on safety and service quality.In the past,managers could only supervise employees’ behaviors through random monitoring or unannounced visits,which wasted both time and energy.With the development of computer vision technology,AI is able to compete with humans’ eyes in terms of recognition capacity.Employing machines to analyze video monitor data in real time can timely and effectively warn persons in charge of abnormal events and help regulate misbehaviors.This thesis focuses on the application of video monitors in the refueling area.Based on deep learning technology,a real-time behavior monitoring system in the refueling area is developed to identify and supervise employee behaviors,improve refueling efficiency and service quality,and finally,reduce potential risks.This thesis mainly completed the following four tasks:First,taking action recognition as the core technology to evaluate action compliance in the refueling areas,this thesis proposes one motion recognition algorithm based on graph convolution and 3D convolution respectively.On one hand,taking the spatiotemporal graph convolutional network(ST-GCN)as the baseline,a new strategy for dividing the adjacent regions of skeleton points is proposed,which increases the adjacent distance in the spatiotemporal dimension and improves the relevance of physically connected joints.Then,based on the salient features of motion,the temporal adaptive network(TAM)and the spatial attention mechanism(SAM)are put forward.Through the amount of motion,frames with large amounts of information and important joints are given more weight,so that the network can learn more useful features.On the other hand,based on the 3D convolutional network,a non-local attention network,ACA-Non-local,is proposed,which solves the problem of insufficient learning ability of action targets for global information and effectively improves the accuracy of action recognition.Finally,the experimental comparison is carried out on the public data set,NTU RGB+D,to verify the effectiveness of points improved before.Second,due to the difficulty in recognizing the action compliance in refueling areas,an action compliance recognition algorithm based on handicraft prior and a dual-branch network action recognition algorithm based on deep learning are proposed.Based on the handicraft prior method and its characteristics,this thesis designs the corresponding logical processing flow to judge the compliance.The dual-branch network contains a high-resolution graph convolutional action recognition branch and a low-resolution 3D convolutional action recognition branch,among which the former is based on skeleton sequence feature,and the latter is based on RGB feature and the target prior so as to propose a person spatial information enhancement network and enhance the spatial relationship characteristics of people.Finally,experiments are carried out on the self-built data set in the refueling area to verify the feasibility of its performance.Third,this thesis analyzes the actual needs of the managers of the refueling areas,applies the action compliance recognition algorithm proposed in this thesis in the actual scene,and develops an intelligent monitoring system for employees’ behaviors in the refueling areas based on Web technology.Considering the difficulty of deep learning technology in large-scale deployment,the system is designed based on the microservice architecture,and the research and development of the system is completed and deployed in practical applications.Fourth,based on the refueling areas,this thesis first collects video monitor data in different periods,different weather and different camera angles,and then makes a picture data set for target recognition and a video data set for action classification,including RGB images and skeleton sequences.