Research On Driver’s Posture Recognition Of Commercial Vehicles Based On Dual-View Video Data

With the development of society,traffic safety is a growing concern of society,while alleviating the severe traffic safety situation will help society develop more harmoniously and steadily.Detecting the driver’s behavior through technical means and taking appropriate forms to intervene in the driving behavior can well reduce the risk of traffic accidents,however,the current research on driving behavior detection can only detects the local driving posture in an ideal environment.So,basing on the requirements of operating truck driver driving behavior detection,build a local database and a light convolutional neural network based on new technology of convolution,to detect driving posture,and through the separation and combination of driving posture to recognition of driving actions.Firstly,Based on previous studies,five types of dangerous driving behaviors are proposed: “Smoking,eating or drinking water”,“Calling”,“One-handed driving”,“Texting or speaking on phone”,“Operating dashboard or navigation”.A questionnaires was designed,and 201 questionnaires was collected,based on the questionnaire data,it showed that there are five types of dangerous driving behaviors generally in the truck group.Also based on 390 segments of video data in a natural driving environment,break the driving behavior into six types of driving actions: “Eatting and so on”,“Taking something”,“Calling”,“Gearing”,“Driving”,and construct physical indicators such as Posture recovery time and Line-of-sight offset time to analysis the risk of driving actions.Secondly,Design a dual-view data collection scheme and collect data after analysised of the data collection environment of the operating truck cockpit.According to the requerements of recognition and the specificity of data,to cut and compress the data.Fiannly,get a SEU-HJSRG dataset containing 109000 images.Considering the recognition models and computing resources,construct a dataset containing 18 categories of 54,000 images,named Data Set-1,and a dataset containing 2 categories of 10,000 images,named Data Set-2.Thirdly,For Line-of-sight detection,builed a small convolutional neural network using only single-view head data,with a model accuracy of 0.99.For Persistent state variable and Transient state variables,build convolutional combination: Block＿A and Block＿B,using network lightweight technology,and build a dual-output convolutional neural network to identify the state variable,with a model accuracy of 0.99 and 0.97.By sharing the shallow features of the first few layers of image,the model reduces the recognition time by 20%compared to recognizing the two perspective pictures separately.Finally,integrate the above models to build a joint recognition model to realize the recognition of driving posture,through 3 examples,the time for the system to recognize a single picture is less than 0.19 seconds,the recognition accuracy is about 0.96,and the model has good real-time and accuracy.Integrated video processing,image preprocessing technology and joint recognition model to build a driving action recognition system,using hyperparameters g and T make the system private.Then take the application in two aspects of streaming media and video detection as an example,discussed the application prospects of the system in driving safety.