Research On Vehicle Anti-drunk Driving System Based On Deep Learning And STM32

In recent years,the incidence of traffic accidents in our country has increased year by year,and the number of people who die in drunk driving accidents is about 100 000 every year.It is imminent to study the prevention of drunk driving.This paper aims to move forward the prevention of drunk driving,drunk driving intervention is mainly achieved by sending instructions such as steady deceleration in advance,the research focus is to explore the limitation and detection methods of drunk driving,and the main work includesBy comparing and analyzing domestic and foreign solutions to the problem of drunk driving,this paper puts forward two feasible solutions: The first scheme is the system built-in blowing pressure and blowing time monitoring program,the driver self breath alcohol detection before starting the vehicle,at the same time face capture program of the system is woke up,and after passing the alcohol test,the face is recorded and the vehicle is started,driving face comparison is used to prevent cheating blowing,this scheme can avoid alcohol driving on the road to the greatest extent.The other scheme is to detect drunk driving based on the driver’s state video,that is,to determine drunk driving by detecting the state characteristics of the facial area,which has the advantages of no blowing and non-invasive.In this paper,a self-help breath alcohol detection machine is designed for scheme 1,and the interference encountered in driving face monitoring is studied.In order to deal with the interference of dark light,uneven illumination distribution,randomness of illumination brightness,and various posture expressions,Gamma transform,facial symmetry repair,random brightness pretreatment training dataset,screening dataset and other pretreatment work are used respectively.To solve the problem of local occlusion feature loss,this paper takes typical mask occlusion as an example and conducts some studies.Based on the Mobile Face Net model,the spatial and channel attention networks are redesigned,and the enhanced feature extraction network integrating the attention mechanism is designed to improve the basic model.The design of the enhanced feature extraction network incorporates the ideas of large residual error,feature pyramid,convolution kernel of different sizes and the granularity of feature extraction.Virtual mask generator is introduced to synthesize virtual mask dataset.Combined with mask shielding characteristics,elastic guidance loss function with attention characteristics is designed to assist model training.In addition,the Ada Cos loss function is used to optimize the basic model loss function.Finally,the upper computer interface is developed based on Py Qt5.In this paper,video dataset including drunk and sober label types are collected for scheme 2,and a pre-processing scheme of face tracking and adaptive scaling is proposed to avoid background interference to the greatest extent.In this paper,three-dimensional convolution is used to "learn" state features,and the key hyperparameters such as extraction frame frequency and frame packet length used in model processing and the improved design of the model are discussed and experimented.Based on the residual network R3 D model,a FR3 D model integrating attention and deep neural network is proposed to realize real-time drunk driving detection,and an integral drunk driving judgment algorithm is designed.