Smoking Detection In Surveillance Scenes Based On Gesture Recognition

With the development of artificial intelligence technology and computer vision is also becoming more and more mature.According to relevant statistics,a large number of smokers will develop major diseases such as lung cancer every year.There are no laws and regulations prohibiting smoking in China,and smoking behaviors are sought after by many young people.People all know that smoking not only harms themselves but also affects the physical and mental health of people around them.Therefore,many public places can only use smoking prohibition signs and artificial supervision.Reminders and human supervision require huge manpower and financial resources and cannot be effectively stopped in real time.Therefore,it is necessary to use monitoring technology to detect smoke-free areas to reduce the waste of more human and material resources.According to the daily smoking gestures,In this paper,the method of video surveillance is used to propose a smoke detection method based on gesture recognition in the monitoring scene,which can better remind the smoking population in time.The main work of this article is as follows:(1)Research and analysis of gesture detection algorithm based on traditional vision gesture recognition algorithm.The traditional gesture detection algorithm is mainly based on the skin color segmentation model based on threshold segmentation and the image segmentation technology based on deep learning.The skin color segmentation technology mainly uses the conversion of different color spaces(for example: RGB->YUV),according to statistics The distribution of skin color in the color space and the sensitivity to the influence of lighting,etc.are used to distinguish gesture targets.Based on the deep learning direction,it is mainly a semantic segmentation algorithm under the framework of a full convolutional network.It infers and classifies each pixel of the image to predict the label and segment different targets;then extract features and classifiers through the convolutional neural network Combine gesture recognition.(2)In the application of smoking detection based on gesture recognition,it is currently faced with the problem that the proportion of the target in the image when the cigarette and the hand are far away from the monitoring device,and the small absolute size of the device itself is difficult to detect the small target.In view of the characteristics of small targets and the characteristics of people's smoking gestures,this paper uses deep learning-based image super-resolution algorithms for research.For current image super-resolution algorithms,image reconstruction is mainly performed by deepening network depth and structural changes.The problem,and the content information required in the recognition task is not all images,so it is proposed to combine the MTCNN face detection algorithm to obtain the image smoking gesture area for image super-resolution reconstruction,which can reduce image reconstruction while increasing the target resolution Time-consuming,reducing the target size of the input network in the later period,can better reduce the calculation pressure of the server and increase the calculation speed.(3)Enhance the local part of the image,use the current mainstream but stage target detection algorithm YOLOv3-tiny for network training,extract gestures and cigarettes in the image,but the accuracy is relatively poor,and obtain more fineness by deepening the network At the same time,the semantic features of s are added to the fusion with the shallow features of the network,and a better detection effect of small-size targets is achieved.Although the calculation time has increased,it can still meet the actual application needs.(4)Introduce and analyze the data set obtained in this paper,and compare the experimental results of the traditional algorithm and the algorithm proposed in this paper under the same data set.The experimental results show that the proposed algorithm has a lot of detection accuracy and the detection speed is obvious It is faster than traditional algorithms and has certain practical application value.