Design And Implementation Of A Smoking Detection System For Operating Vehicle Drivers Under Surveillance Video

With the increasing demand for the quality of travel,uncivilized driving behavior has become the focus of regulation and rectification of the entire industry.In which the most common is the driver smoking behavior,this behavior not only harm the health of passengers,but also easy to cause traffic accidents.So it is very important to detect the smoking behavior of drivers.At present,the illegal behavior of drivers smoking is mainly through the manual view to monitor the video way,easy to appear fatigue and human law enforcement problems.Therefore,the development of smoking behavior automatic detection system can reduce the waste of labor and time and improve the efficiency and quality of the inspection.However,most of the current smoking behavior detection algorithms are focused on smoking action or smoke characteristics,so when the driving environment changes,it is easy to cause misjudgments and omissions.In response to the above problems,this paper proposes an optimized system for detecting smoking behavior of drivers of operational vehicles under in-vehicle surveillance video.The main work of this paper is as follows:(1)Establish smoking testing data sets.At present,there is no common smoking behavior data set,so the data in this paper is mainly the video shot by the real car scene.During the shooting,the cameras were mounted under the front mirror of the car and on the driver's dashboard.Since the diversity of samples can increase the validity and robustness of the depth model,the self-recorded videos cover different weather,lighting conditions,genders and smoking positions.At the same time from the network to collect some smoking scene video and pictures to enrich the data set.Then,the labeling of the type and position of the data set is processed.In this dataset,there are nearly 100,000 frames and 5308 manual annotations,which are suitable for training and experiment.(2)In view of the small range,thin color and low concentration of smoke produced by smoking compared with the smoke produced by general combustion,a small image smoke detection method based on feature fusion is proposed.In order to make the model study the feature of smog area better,this paper divides the labeled smog area and non-smog area into16 * 16 overlapped small range images,Its LBP and HOG features are extracted to train a support vector machine(SVM)to find the best classification hyperplane between smoke and non-smoke regions.The interference caused by thin smoke and low image definition is effectively optimized.During the detection process,the LBP and HOG features are extracted from the suspected smoke regions based on the hybrid Gaussian model and the frame difference method respectively,and the features are fused,then put into the trained support vector machine(SVM)to mark the real smoke region.(3)In order to solve a series of problems such as large resource consumption and high time cost,an improved cigarette detection method based on YOLO v5 s is proposed,in which Bottleneck and standard convolution in Dark Net are replaced by lighter Ghost Net and depth separable convolution to optimize the calculation consumption of the model.In addition,according to the characteristics of vehicular video data,only one feature layer is output in YOLO V5 S,which effectively avoids the detection interference caused by small target learning.(4)Combining the front-end page display technology and the trained detection algorithm model,a detection system for drivers' smoking behaviors is constructed on the Web platform for the operation and use of relevant persons.