Distracted Driving Behavior Detection Based On Lightweight Convolutional Neural Network And Embedded Implementation

With the increasing number of cars,traffic safety has increasingly become an issue of great concern to the society.At the same time,as more and more distracted sources enter the driving process,drivers’ line of sight deviation or inattention behavior occurs frequently,which can easily lead to traffic accidents,so it is of great significance to actively detect distracted driving behavior.At present,the research on distracted driving aiming at the visual characteristics of drivers is focused on the use of intrusive and expensive infrared equipment,and there is a lack of detection methods and indicators based on deep learning.However,convolution neural network has become an effective means of visual analysis.In view of the transient and frequent characteristics of distracted driving behavior,this thesis focuses on visual distraction and proposes a lightweight distracted driving recognition algorithm based on deep learning technology.The research content includes the detection method of driver’s facial features and posture angle,the construction principle of algorithm model and the establishment strategy of distraction judgment index.First of all,a lightweight face detection algorithm based on Blaze Face is proposed.The face region is anchored by cascading multi-scale features,and the eye ROI region is segmented by facial "three courts and five eyes" distribution features,so that facial and eye features can be extracted quickly and accurately.Then,the line-of-sight estimation algorithm of multi-stage fusion of head posture and eye movement angle is studied,and the head posture and eye movement angle estimation model SDA-Net,is proposed,and the algorithm theory,network structure,mixed attention mechanism and compensation mechanism of the model are described in detail.Finally,the distracted driving state recognition algorithm based on Atten D is studied.According to the gaze angle estimation after fusion,whether the driver’s attention deviates from the road is inferred,and a sliding window is designed to detect the video frame.The thesis has carried on the full experimental analysis to the above research content,and has carried on the embedded test.The lightweight face detection algorithm based on Blaze Face is verified in the public data set WIDER FACE,and the proposed algorithm achieves the accuracy of 91.20%,89.6% and 84.30% in Easy,Medium and Hard,respectively.The MAE of SDA-Net with mixed attention mechanism is 4.62 on the head posture data set AFLW and 4.69 on the eye movement estimation data set MPIIGAZE.The effect of the model is obviously better than that of the traditional convolution neural network.On the collected driving data set,the accuracy of judging normal driving and distracted driving is 97.2% and 96.54% respectively.The results show that the distracted driving detection algorithm in this thesis has high accuracy and robustness,and can accurately identify distracted behavior.