Research And Application Of Cigarette Detection Technology Based On Deep Learning

It is well known that smoking is harmful to health.Smoking in public places is not only related to smokers themselves,but also forms second-hand smoke to endanger the health of other people around them.Moreover,fires in public places will be induced by irregular smoking behaviors,which will cause inestimable losses to all sectors of society.Therefore,increasing smoking bans in public places has become an irreversible trend.However,the detection effect of the technology of current smoking detection based on smoke sensors is not ideal due to the interference of environmental factors such as ventilation,and the detection accuracy of the smoking detection based on traditional computer vision is not high.Therefore,it becomes more and more urgent to research how to quickly and accurately detect the behavior of smoking.With the rise of deep learning,the technology of cigarette detection based on convolutional neural networks(CNN)has been used to identify smoking behaviors and has achieved certain results.However,as the main researched method at present,the performance of cigarette detection based on CNN still needs to be further improved due to the influence of internal and external factors such as smoking posture,illumination,and complex backgrounds in actual scenes.Therefore,this paper mainly researches the technology of cigarette detection based on CNN to improves the existing defects in the current cigarette detection algorithm so that the accuracy of recognition is improved.The main research contents include:(1)The target detection algorithm based on CNN detects the frame by frame of video.Due to the shortcomings of this type of algorithm such as complex structure and many parameters,the overall cigarette detection is too complicated,which makes the detection time too long,and the computer hardware occupation rate is too high.Therefore,a rapid cigarette detection algorithm based on faster region-based convolution neural networks(Faster RCNN)is designed in this paper.The main ideas are: 1)Detecting human faces and the pictures of face are used to as areas of cigarette detection,thereby the area of target detection and the false detection rate are reduced effectively;2)The method of color segmentation is used for the initial detection of cigarettes in the face area.If there is a high probability of cigarettes,the Faster RCNN algorithm is used for detailed detection of cigarettes.If the probability of cigarettes is low,the current video frame is not processed.As a result,the number of Faster RCNN algorithm runs is greatly reduced,which makes the detection efficiency is improved and the hardware occupancy rate is reduced effectively.The experimental results show that the hardware occupancy rate and detection time of the rapid cigarette detection algorithm based on Faster RCNN is reduced significantly on the premise of ensuring the detection accuracy,and the rate of false detection is reduced by about 2% on the original basis.(2)Cigarettes are small targets in video frame images,and the detailed feature information of small targets in CNN deep feature maps is often too small or completely lost.Therefore,this paper focuses on the feature pyramid network(FPN)to improve the existing feature extraction methods,and designs a method for cigarette feature extraction based on feature fusion.This method fuses the shallow local features of the CNN with the deep global features,thereby the problem of the loss of small target features in the forward propagation process is avoided effectively.And the method is applied to the forward network of Faster RCNN for cigarette detection.The experimental results show that the improved Faster RCNN cigarette detection algorithm based on feature fusion can reduce the missed detection rate of cigarette detection effectively and improve detection efficiency on difficult test sets.(3)Combining the above feature fusion method with rapid cigarette detection algorithm based on Faster RCNN,an improved Faster RCNN rapid cigarette detection algorithm based on feature fusion is designed.Compared with the above two algorithms,the combined algorithm can combine the advantages of the above two improved strategies,thereby the false detection rate and missed detection rate of target detection are reduced simultaneously on the same data set,and achieve better detection efficiency.(4)In order to verify the practicability of the improved Faster RCNN rapid cigarette detection algorithm based on feature fusion,an intelligent smoking behavior detection system is built in this paper.The system consist of three modules: face detection and recognition,cigarette detection,and smoking information recording.The proposed algorithm is applied practically.