Research On Detection Algorithm Of Driver Fatigue And Distraction Behavior

In recent years,the number of motor vehicles in China has been increasing,and the increasing number of drivers has led to the increasing probability of traffic problems year by year.According to relevant research,the non-standard operation behavior of drivers and fatigue driving are the main causes of accidents.Therefore,timely monitoring and early warning of these two behaviors become more and more important.The existing detection methods are difficult to meet the requirements of accuracy and speed at the same time,so the research contents of this paper are as follows.For fatigue detection,starting with the research of traditional facial behavior and detection algorithm based on deep learning,aiming at the imbalance between the accuracy and real-time of the current algorithm,a fatigue driving detection algorithm based on MF-YOLOv4 is proposed,which transforms fatigue detection into target detection,and finally detects the driver’s state according to the fatigue judgment index.The specific improvements are as follows.Firstly,Mobile Netv2 is used to replace the original CSPDarknet53 backbone extraction network,and the joint convolution channel parameter alpha is reduced to reduce the number of layers of the whole network.Secondly,the FPN tiny lightweight feature pyramid module is added to filter the redundant information in the image.Finally,aiming at the detection target volume in this paper,the original soft NMS of YOLOv4 is improved.In the detection process,there is no need to consider the score and coincidence degree of the target frame at the same time to get the results,which further improves the detection speed of this method.In this paper,the fatigue detection algorithm is tested on the self built fatigue detection data set.The overall average precision(map)is 98.37%,and the speed is increased to 46 frame rate / s compared with that before the change,which meets the needs of real-time detection.The volume of the algorithm model is only 49 MB,which is 197 mb less than that before the change.A target detection method based on SC-YOLOv5 s convolutional neural network is proposed for the distracted actions of drivers(smoking,calling and drinking water).The specific improvements are as follows.Firstly,by deleting the original network large object detection scale,the model can better meet the target to be detected.Secondly,the attention module of Se channel is introduced to retain the key information of the picture,which can improve the detection speed with a small number of parameters.Finally,Dlou function is introduced to directly minimize the distance between two target frames a and B,which solves the problem that Glou loss function consumes a lot of time and contacts the prediction frame with the real frame,which affects the convergence speed.The average accuracy of the model in this paper is 99.2% and the speed is 58 fps on the three behavior data sets,which is 1.4% higher than that of the unmodified YOLOv5 l model,and 12.3% higher than that of SSD algorithm.The algorithm model in this paper not only takes into account the speed,but also maintains high accuracy.