Research On Smoking Action Recognition Based On Skeleton Information

Smoking harms the health of smokers and second-hand smokers.Fires caused by smoking are not uncommon.With the extent of public concern about their own health and public safety have become increasingly demanding,smoke detection has become a hot topic in academia.The existing smoking detection has the following problems.The detection cost based on the sensor is high and the accuracy is low.Wi-Fi signal-based detection has unstable performance in noisy environments,and video image-based detection has problems such as high false detection rate and low real-time performance.Aiming at the problems of smoking detection accuracy and real-time performance,this thesis proposes a periodic smoking action recognition model and a spatial residual hybrid graph convolution model by analyzing the characteristics of smoking action.The main work of this thesis is as follows:(1)Aiming at the problem of low accuracy of smoking detection and recognition based on video images,this thesis constructs a stable performance recognition model for smoking action.We manually add non-skeletal connections in human body modeling,set different weight matrices to distinguish different connections,strengthen the information exchange between nodes,and make the feature information learned by the deep network more stable and accurate.The inaccurate recognition problem caused by the information mixing caused by the superposition of different domain convolution orders.Secondly,aiming at the problem of overlapping joint points during smoking,this thesis uses skeleton edge graph convolution to better capture the dependency between bones and reduce the loss of high-dimensional feature information.The length and direction of the skeletons can provide more effective features,and the main parts of the skeletons do not overlap completely.We design a hybrid model of joint graph convolution and skeleton edge graph convolution,which combines the advantages of these two models.The smoking action recognition accuracy reaches 93.4%.(2)Most of the existing research on smoking action recognition is only for single person,and the problem of poor real-time performance of multi-person action recognition.This thesis proposes a multi-person smoking action recognition method based on joint point detection in a multi-person scene.The thesis analyzes the rule of smoking action,conducts a large number of experiments,studies the regularity of time and action trajectory during the whole smoking process,and standardizes smoking actions.Experimental results show that the recognition accuracy of smoking actions for multiple people can be kept stable in different environments.In summary,this thesis designs a multi-person smoking action recognition system.It can recognize local video and real-time shooting video,and can realize real-time smoking behavior detection of multiple people through a defined interface,and has a high accuracy.The smoking action recognition model proposed in this thesis is subjected to a large number of experiments on the data set.It compares with existing motion recognition models.The experimental results show that the performance of the model proposed in this thesis is stable,and it provides scientific and effective application value for smoking detection in public places in the future.