Research On Driving State Detection Based On Image Recognition

With the economic growth,China’s car ownership is increasing year by year,which leads to a lot of road safety problems.Fatigue driving and distracted driving are the two main causes of traffic accidents.It is of great significance to study the driver’s driving state and give early warning when necessary to ensure road safety.The research of driving state based on machine learning is mainly to detect and analyze the features of driver’s face,head or hand through images.The method based on image technology has the characteristics of cognitive similarity,non-invasive,strong robustness and so on.It has great development potential and application prospect in the research of driving state.In the complex driving environment,there are higher requirements for the speed and accuracy of model detection,and the problems of illumination and feature occlusion can not be well solved.This paper mainly studies the driving state from the following two aspects.Firstly,the driver’s fatigue state is mainly studied from the eye features and mouth features.In order to improve the diversity of face pose and size in face detection,lfpw,and AFW data sets with key points annotation are used to construct fatigue detection data sets.The three data sets are upper body images in complex environment.The face size,pose and background of the data are very rich,which can effectively improve the robustness of the model.In order to adapt to the feature extraction of face region in real driving environment,yaw-dd fatigue driving video is used to expand the data set,and CEW eye closure data is used to expand the eye closure feature.Secondly,the network structure is adjusted according to the statistical characteristics of the data set to make the model more compact.Then,the YOLOv3 model is used to detect the features of face,eyes and mouth.The driver’s fatigue state is determined according to the number of eyes closed and mouth yawn frames.The attention mechanism is introduced into the detected face area to make the detected back frame search the features according to the previous face area.The detection speed is maintained while the eye and mouth small eyes are improved The ability to check the target features.Finally,in order to speed up the network reasoning,channel pruning technology is used to compress the model,which effectively improves the speed of model reasoning and effectively reduces the model parameters.Secondly,for the study of distracted driving state,mainly from the feature extraction and analysis of its mobile phone.Because the mobile phone is the main factor causing the distracted driving,then the water cup is introduced as another distracted driving feature,which is used as the interference to the mobile phone feature,so that the robustness of the model is higher;the two hands holding the steering wheel is used as the normal driving feature.Then,the YOLOv3 model is improved to extract the three features for a single regional feature.Finally,the distracted driving state is judged according to the maximum confidence of detection.The regional feature extraction method used in the model introduces the position information of the feature through the label,which is more effective than other single label classification methods based on the whole image feature,with higher accuracy and robustness.Finally,the pruning technology is used to compress the model,and the detection speed and model size of the algorithm are greatly improved.