Research On Driver Distraction Recognition Method Based On Deep Learning

The distracting behavior during driving is frequent and transient,and it is easy to cause road safety accidents.Monitoring the driving end and warning in time is an effective means to reduce the risk of collision.This article selects the head posture that can reflect the driving state as the main body of research,and uses deep learning theory and methods to realize the recognition of the driver’s distraction state.The main research contents are as follows:First,a head pose estimation method based on deep learning is proposed.In view of the unfavorable factors existing in the actual image acquisition process,such as complex light source,great change of light and shade and vehicle bumps and vibrations,the training set is enhanced to simulate the real driving scene before the experiment.Through downsampling,Gaussian blur and boundary enhancement filtering and introducing randomness in the process,the model is forced to learn pictures of different image quality and reduce the dependence on high-definition images.According to the two design ideas of single regression and classification combined with regression,AlexNet and ResNet,are improved to give five convolution networks with different depths.In the idea of classification and regression,the decoupling training of Euler angle is carried out,the corresponding compound loss function is set and the training strategy of transfer learning is used.finally,the accuracy of the two ideas is verified on each model at the same time.The experimental results show that the effect of classification combined with regression idea is more prominent.The optimal model HPE₁01 in this paper can achieve an average absolute error of 5.61°on the data set AFLW2000 and an average precision of 79.2%（±3°）and 92.3%（±5°）on the data set AFW,and has a certain anti-jamming ability,which is suitable for the task of head pose estimation in fuzzy environment.Second,the head frame is marked and the difference is analyzed.Different from the public data set,there is no annotation information on the driver image in reality.In order to more accurately analyze the relationship between head posture and distracted driving,this paper uses the deep learning model RetinaFace to detect the face position,and then enlarges the boundary box by expanding the coefficient to increase the proportion of the head.Euler angle data for each image in the public distracted driving dataset SF3D（State Farm Distracted Driver Detection）were calculated using the model HPE₁01 after completing the annotation,and the differences between classes are discussed qualitatively and quantitatively.The analysis of variance shows that there is a statistically significant difference in head posture between safe driving and all kinds of distracted driving at 95% and 90% confidence levels,and the postures of all kinds of driving movements are distributed in a specific Euler angle range,which provides a characteristic basis for the design of subsequent recognition methods.Finally,a distraction recognition method based on Euler angle of continuous video frames is proposed.A single frame image cannot express the continuous information in a period of time,so it will produce a great chance to distinguish the distracted state,so the method can fully express all kinds of head movements in driving from the video of the driver.Based on the distracted driving theory,three kinds of distraction parameters are set,the dimension reduction is completed by calculating the Euclidean distance,and the corresponding safe driving head range and distraction threshold are obtained according to the current imaging angle statistics.The experimental results show that the driving state can be clearly distinguished by calculating the distraction parameters of the video frame to be tested,and it has a good filter for the driver to observe the normal driving operation of the rearview mirrors on both sides.The paper has 38 pictures,23 tables,and 76 references.