Research On Fatigue And Distracted Driving Detection Method Based On Deep Learning

In recent years,the number of vehicles has increased,and road traffic accidents have been increasing year by year,posing a threat to the safety of life and property.The investigation found that driver fatigue and distraction are the main causes of road traffic accidents.With the rapid development of computer vision technology and the continuous iteration of models,computer vision technology is widely used to solve various problems.Computer vision based detection methods have become the mainstream method for fatigue and distracted driving detection due to their high accuracy and low deployment difficulty.This paper mainly studies the detection method of fatigue driving and distracted driving based on computer vision technology,and formulates relevant improvement schemes for the problems existing in the detection model in terms of real-time and accuracy.The target detection method based on YOLOv5 is applied to the driver’s fatigue state detection,and the ME-YOLOv5 fatigue driving detection model is proposed.The specific improvement plan has the following points: Combined with the size characteristics of the target to be detected,optimize the detection specification and scale of the model,and make the model more inclined to the detection of small targets by removing the detection layer for large targets;in order to speed up the feature extraction speed,The lightweight network Mobile Netv3 is used for feature extraction;in order to make up for the defects of the backbone network’s receptive field and enhance the network’s ability to extract important features,the SAM is introduced to broaden the feature capture range and complete feature recalibration;the Bi FPN structure is introduced,Complete the adaptive fusion of multi-level feature maps to enrich the feature expression.The ME-YOLOv5 detection model achieved an average accuracy of 95.85% on the self-made data set,and the detection speed was 132 FPS.For distracted driving detection,this paper will build a Res Net50-PSAM distracted driving detection model based on the Res Net50 image classification method to identify nine typical distracted behaviors such as drinking water and answering the phone.An improvement plan is proposed for both classification accuracy and detection speed.The specific improvements are as follows: the driver’s hand and face areas contain important features that are helpful for correct classification.The SAM is introduced and its structure is adjusted,divide the feature map into several blocks,obtain the weight representing the importance of each block,and amplify the differences between blocks through the Tanh activation function,so that the entire area of the hand and head receives more attention.In order to balance the proportion of losses generated by samples,the cross-entropy loss function is adjusted to reduce the loss of samples that are easy to classify and mislabeled samples.In order to meet real-time requirements,structural pruning is used to complete the lightweight model.Finally,Res Net50-PSAM achieved a classification accuracy of 95.01%on the State Farm dataset,reduced the number of parameters by 25% on the basis of Res Net50,and accelerated the detection speed by 14 FPS.This paper builds a fatigue driving and distracted driving detection system,combines the ME-YOLOv5 fatigue driving detection model and the Res Net50-PSAM distracted driving detection model to realize the visualization of detection and results,and complete fatigue in the real cab The simulation test of driving detection and distracted driving detection further verifies the practicability of the model.