Research And System Design Of Smoking And Calling Action Recognition Algorithm Based On Deep Learning

Smoking and calling are two common behaviors in our life.If cigarettes or telephones are used illegally in certain places,they can cause serious safety threats.In recent years,due to the rapid development of deep learning technology in the field of computer vision,when the camera is facing the face,the action recognition of smoking and calling has achieved good results,but the effect is still not ideal for complex scenes in practical application.The main reason is that the camera has different angles in different scenes,leading to the large-scale transformation of the characters in the image.When the person is far away from the camera,the two kinds of objects,namely cigarette and telephone,are usually small objects with no obvious visual characteristics.Furthermore,the actions of smoking and calling may occur at the same time and there are different rules or tolerances for different actions in different scenes.However,most of the current methods regard the action recognition problem as the image classification problem and only distinguish the two categories of "smoking" and "calling",which is difficult to meet the actual application requirements.Above all,the action recognition of smoking and calling is divided into two stages: suspected object detection and action interaction recognition.The main work and results are summarized as follows:(1)In the stage of suspected object location,to improve the accuracy of small object detection,a two-way pyramid network with spatial information enhancement is proposed for feature extraction.The image pyramid is used to improve the network’s ability to represent texture details,and a spatial information enhancement module is proposed to reduce the interference of feature maps in redundant or irrelevant regions.In the classification and regression of candidate box instances,using the interaction between people and objects in smoking and calling as prior knowledge,a structural clustering guidance module is proposed.This module suppresses the candidate frame relation of non-smoking and non-calling and further improves the object detection accuracy of human interaction with cigarettes or telephones.(2)To avoid missing detection when only using an object detection algorithm in the case of serious occlusion or blurred image,a reconfirmation algorithm for suspected objects is proposed.The human key points are used to estimate the human posture and judge the suspected smoking or calling posture,thus filling the gap in object detection algorithm results.At the same time,the camera is further controlled to focus and magnify the location of the suspected object,creating good image conditions for the second-stage action interaction recognition.(3)To make action recognition adaptable to the simultaneous occurrence of smoking and calling,and can be flexibly applied to scenarios with different rules or tolerances,an action interaction recognition algorithm based on visual features and spatial key points features is proposed.Human-object visual features and spatial key points features are generated to subdivide the action of smoking and calling using the existing object detection and key points detection results.At the same time,the environment-related features are proposed by using the visual information of people and their surrounding environment to capture the smoke features in the environment and further improve the accuracy of smoking action recognition.(4)Combined with the proposed algorithm,an intelligent monitoring system for smoking and calling action recognition is developed.For a network input stream,the system recognizes the interaction of smoking and calling in real-time through the strategies of keyframes reasoning,load balance,and smooth prediction of intermediate frame,and visualizes the results on the interface,which provides an intelligent supervision basis for the management department.