Research On The Application Of Cloud-based Compensation Detection Method In Driver Smoking Behavior Detection

In recent years,China has been actively using information technology in the work of transportation safety,constantly strengthening the "two passengers and a dangerous" key areas of supervision,strengthening the source management of road transport enterprises,while extending the control of the behavior of vehicles and drivers.Road transport enterprises use intelligent video monitoring terminals and vehicle dynamic monitoring platform to supervise and manage daily transport tasks,of which unsafe driving behavior monitoring and early warning is of great concern.In the actual work in the field of intelligent transportation,it is found that there are false alarms from vehicle terminals to driver behavior alarms,especially the false alarm rate of driver smoking behavior is high,resulting in too much such false information in the enterprise platform,which not only causes misjudgment to the enterprise monitoring and warning,causes interference to the driver,but also causes trouble to the line management department assessment enterprise.Therefore,this paper proposes and studies the cloud compensation detection method between terminal and platform for the problem of too much false information of driver’s smoking behavior in the enterprise platform,and uses convolutional neural network to review the alarm data.Through experimental comparison and model selection,it proves that the Faster R-CNN algorithm model is applied to the platform active safety warning subsystem as the algorithm model of cloud compensation detection system,which can effectively detect and screen out terminal false alarm data and reduce the platform false alarm rate,and this paper improves on the traditional Faster R-CNN algorithm model to achieve more accurate identification of driver smoking behavior.The main work of this paper has the following aspects:(1)The driver smoking behavior target detection algorithm is studied based on the real measurement data generated by terminals in the enterprise platform.To improve the performance of the network model,a large amount of real in-vehicle terminal alarm data,including driver smoking data and unaffected smoke images captured by DSM cameras in the cockpit and ordinary RGB cameras in the room under different scenarios,are collected from the enterprise,composed into a dataset and randomly assigned.In order to maintain the basic features of the images,the original dataset is expanded by rotating,blurring,adding noise and adjusting brightness in various ways during data pre-processing.In this paper,SSD,YOLO V5,and Faster R-CNN algorithms,which are commonly used in engineering,are selected to train the model,validate and evaluate the test set.The measured data show that SSD has lower performance,Yolov5 has the fastest convergence and detection speed,and Faster R-CNN has the best accuracy.Considering that the scenario proposed in this paper requires algorithms with higher detection accuracy,the Faster R-CNN model is selected for target detection of alarm data.(2)Based on the Faster R-CNN algorithm model,after analyzing its advantages and limitations,the algorithm is improved in two aspects.First,the Softer NMS algorithm is used to optimize the non-maximal suppression(NMS)to improve the candidate frame accuracy,avoid losing important detection frames,and enhance the category confidence;second,the CIo U loss function is introduced to better handle the target frame overlap and shape change,which in turn improves the detection accuracy.Experimental results show that these improvements improve the detection accuracy and speed,and achieve more accurate driver smoking behavior recognition.(3)A cloud-based compensation detection method is proposed based on the problem of excessive driver smoking behavior false alarm data in the enterprise platform.The research designs a cloud-based compensation detection system for driver smoking behavior in enterprise vehicles and applies the system to the enterprise vehicle dynamic monitoring platform.The Faster R-CNN algorithm model trained in the first two works is used in the cloud to detect the target and screen out the false alarm data generated by the terminal,and the detection results are used as the platform alarm information.The test results show that the algorithm model can effectively detect the false alarm data transmitted by the terminal,while the cloud-based compensation detection system can effectively screen out such false alarm data and significantly reduce the false alarm data of driver smoking behavior in the enterprise monitoring platform.The research results of this paper improve the accuracy and evidence validity of driver smoking behavior alarm information in the enterprise vehicle dynamic monitoring platform,reduce the interference caused by a large amount of false alarm data in the enterprise platform,provide stronger technical support for problem retrieval,accident reversal,and enterprise supervision,and also provide more accurate assessment data and decision basis for the industry supervisory department,thus reducing the road transportation accident rate,avoiding Prevention of major transport accidents,to protect road transport safety production.In addition,the detection data of the cloud-based detection system can also provide data support for the optimization of DSM algorithm of vehicle terminals.